general

[winter geyser]

Why doesn’t the DeepSeek R1 0528 model on Web LMArena output any code? Is it a bug, or is code generation disabled for this model?

[ocean vortex]

I feel like people should stop pushing the narrative that new R1 is equivalent to o3...

HLE: 20.6 vs 17.7
SimpleQA: 49.4 vs 27.8
SWE Verified: 69.1 vs 57.6
(o3 vs R1)

it's a great model but it is not quite on the level of o3 or 2.5Pro

[sturdy mica]

https://cdn.discordapp.com/attachments/1170641241193586769/1373958744278306867/IMG_0213.gif

hi chat

[brittle hull]

Yo, has anyone seen a model codenamed Stephen on lmarena today? I came across it twice

[patent aspen]

deepseek model

probably r1

[hardy pecan]

Is it good?

[brittle hull]

Maybe pre-Grok-3.5, 'cause Musk dropped early Grok3 a week before last time

[brittle hull]

Over good

[patent aspen]

No it's Chinese

I think it even says deepseek in the code

[brittle hull]

Unlikely R2, since testing R2 right after releasing Updated R1.1 is kinda meh

[alpine coral]

oh..

[sacred plaza]

Are there any grok 3 glazzers on this chat. If so, what is your best steelman argument for why I should use anything from xAI given the fact the Twitter guy brainwashes grok's system prompt on a monthly basis.

[tall summit]

it's free

with limited uses

[fringe carbon]

i use deepseek. it is very good at password management. you can give it all your passwords and not have to remember them again

[tall summit]

LMAO

[fringe carbon]

very good context window

ok real talk though

is it safe to upload stuff like api keys and whatnot to openai

probably right?

i'm not afraid openai gonna use them

i'm afraid that the model spits out the key to someone else

[vernal meadow]

@sacred plaza doesn't sound like you are asking for any usecases. Sounds like you are already set.

@fringe carbon You should have a paylimit on all API keys and from time to time use new ones. These days the API providers make it quite easy.

I woudn't post my passwords in there tho. Not sure in which context that would make sense : P

[fringe carbon]

well i write super spaghetti code

and hard code my api keys at the top of files

so in that context

[ocean vortex]

chances of that happening even if they fail sanitizing data and train the model on your exact chat are actually incredibly slim. It's very unlikely to output your exact key without changing a single char

bluntly speaking it will get lost in the sea of data. And since they are not gonna overfit the model on your chats there's no chance. Not to mention that them not sanitizing your data is now what you would expect either...

or in other words still... a key is not a thing that is easy to "remember" unless it is being shown repeatedly/overfit. It knows generally how it looks but not this specific key exactly as it is

[unborn ocean]

@fringe carbon you can just disagree to data training (in the settings for chatgpt / gemini) and they are legally not allowed to train on the stuff

that is how i do it

and with aistudio I am just extra careful

[sacred plaza]

no, i want to hear the best argument why people are using it. not going to change my mind on using it since i am not on twitter. willing to change my mind of its usefulness in terms of capability though. access to real time twitter data does seem like a moat

[dusky aurora]

real time twitter data is an oxymoron

[sacred plaza]

close to real time twitter data****

[echo aurora]

reminder we're going to be watching this in a bit for anyone that wants to join! https://discord.gg/Vk7QXKXf?event=1377683812024189068

[sacred plaza]

thanks for sharing! will tune in as well

[misty vault]

https://tenor.com/view/cat-look-cat-look-at-camera-silly-cat-in-a-cage-gif-889392959852579879

[dusky aurora]

like I said, twitter is a thing of the past. real time x data is possible, but twitter not

[bright lion]

Oh Boy, can‘t wait to pay over 200$ a month for restricted o3 pro access now

Where did you hear this from btw.

[misty vault]

twitter

[echo aurora]

Starting the a16z podcast now!

come join #1340554757827461215

[misty vault]

what is ur opinion on gpt-4-0314-32k

[wintry tinsel]

Thoughts on the new deep seek for festive writing?

Is V3 or R1 V2 better?

[elder rapids]

what is this server

[tall summit]

how do you not know

[keen beacon]

so goldmane is gonna be ga 2.5 pro it would seem

[elder rapids]

ye

thank God

y'all kept saying redsword was better

I was like youre tripping

[keen beacon]

i just need it in aistudio + raw thoughts 🤩

[elder rapids]

ong

man

give me that model

😭

when do you guys think it'll be released

GA taking too long bruh

[keen beacon]

next month

[elder rapids]

no shi

[teal mantle]

anyone guess what model made this

[keen beacon]

🤣

[elder rapids]

I mean like when next month

[teal mantle]

it is so far one of the worst ones

[keen beacon]

probably soon enough since they removed redsword

[elder rapids]

random model

send in dms

the invite

[kind cloud]

ok

[teal mantle]

nah it is o3

it sucks at this (once)

[elder rapids]

0% formatting

[teal mantle]

it still sucks and don't know what is a Touhou spellcard

[elder rapids]

goldmane codes beautifully

someone gotta talk about that

did you guys see redsword and goldmane code? 😭

ts is a work of art

the gradient indents are so beautiful

[kind cloud]

I think we can distinguish between Gemini-flash and Gemini-pro because Pro accurately remembers chapter titles of One Piece, but Flash doesn't. As a result of this knowledge test, goldmane is identified as Gemini-pro.

[elder rapids]

we do know for a fact it's Gemini 2.5 pro tho there's no need to test it

just waiting for Logan to say "Gemini"

[keen beacon]

yeah

[elder rapids]

flash has sauce

[keen beacon]

but it does know less though

something happened between flash and flash lite

i noticed

[elder rapids]

wonder how they're going to scale up the diffusion model

what

that's an unnecessary distinction here

more loads, larger model, maintaining efficiency

yeah well I don't

[keen beacon]

im curious about this. if u can answer this, do you see it ever replacing gemini/being the main thing? (diffusion)

[elder rapids]

Google prolly doesn't see it as replacing, just a new strong route

but I don't believe they dont know what theyre going to do with it

[keen beacon]

yeah, but id like to know their opinion on it

[elder rapids]

I would think theyre planning to integrate it with search

or large text updates

and not necessarily discrete generation

[kind cloud]

This means that if it answers 'I am a large language model, trained by Google.' and correctly answers the chapter title, it is 'goldmane' in almost all cases. This is solely how to continue longer chats with goldmane on new lmarena, as far as I know.

[teal mantle]

I forgot but is o3 parameter identical to non-reasoners like 4o?

[keen beacon]

yes

[misty vault]

brian is google ceo???

brian is part of that team???

me too

[golden ocean]

same

[misty vault]

sydney chatbot

Im not interested in building ai if i cant build a gpt-4-0314 at home

I will need datacenter

[elder rapids]

give a single example

I want to use it

like, the specific chapter

[kind cloud]

chapter1117 'mo'

[elder rapids]

@balmy mist flowith corrupted some of my stuff

[hollow ocean]

Today confirmed

✅

Maybe

[tall summit]

based

[keen beacon]

adhd is a scam to get kids on meth

the medicine they give kids with adhd

is brain rotting

my brother went schizophrenic off it

im telling you rn dont take adderal

the difference is adderal is addictive

wut lol

you really shouldnt rely on adderal to function

you should look for other remedies

teas and stuff

ai is getting really good at study guides as well

look up NotebookLM

alot of people take adderal for school

i mean around me atleast

there are alot of kids who only take the medicine to focus in school

thats a symptom of a larger problem tbh

trash courses

our schooling is just trash

wym

[keen fulcrum]

Have you seen Perplexity Labs?

https://www.rxddit.com/r/perplexity_ai/comments/1kypixi/introducing_perplexity_labs/

[keen beacon]

oh yea for sure

but as a first generation US citizen imma tell u one thing

relying on anything to function is bad

that was my point

not that it's not a real disorder

its just under researched

i was watching a video from neil de grass tyson, and yk how the common saying is "we only use 20% of our brain" etc

thats not the truth

we only know what 20% of it does

ain a flex bro 😭

liver gon be fried by the time u older

dont hate the messanger

[late path]

A person with mild ADHD would also live a better life if they took amphetamines, I guess?

[keen beacon]

no

they're trying to use AI to create medicines genetically tailored to ur dna

so instead of getting pills with side effects you'd have drugs tailored for you

https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2023.1335901/full

https://www.pfizer.com/news/articles/artificial_intelligence_on_a_mission_to_make_clinical_drug_development_faster_and_smarter

welcome to the future lil bro

for sure

US government is tryna use it to end HIV

etc

wow it got auto modded

for sure

[misty vault]

crispr

who is deleting their messages

[keen beacon]

idk doubt it

it can either go 2 ways

ai helps us or ai destroys us

[misty vault]

yeah we got gpt-4o

[keen beacon]

https://www.youtube.com/shorts/5Yh9U5tAuO0

thats just the ai we have now

companies 100% have gatekept ais

like google has a self learning huge new ai

[elder rapids]

goldmane is so intelligent this is crazy

😭😭

nebula moment

istg bro

[elder rapids]

I'm being so deadass

[keen beacon]

lol

[elder rapids]

thinks quickly

[keen beacon]

i doubt its deepthink

[elder rapids]

ye

[keen beacon]

we know what it is already like we keep arguing about this lol

[elder rapids]

ong

seriously though

wild,

goldmane is actually so good

it's so smart

nah like actually

I know it's some shi to compare the models

I love playing with o3

but goldmane is GOOD

[keen beacon]

did u see this btw? 😭 https://discuss.ai.google.dev/t/massive-regression-detailed-gemini-thinking-process-vanished-from-ai-studio/83916/84

rip raw thoughts

[elder rapids]

nah it's still up in the air

[keen beacon]

On summaries, we have heard a lot of valid feedback. We understand this is a different experience from the raw thoughts previously available in AI Studio. Sometimes product teams have to weigh a lot of pros and cons to come to a specific decision. This is one of those times. Please work with us and help us in getting summaries to a point where they have just the right amount of detail that you need. You are our valued and needed collaborators in this. In the meantime, we will keep listening to your feedback here or DM @shresbm or @vish_owl or @OfficialLoganK on X.

based on that it basically confirms its a competitive decision i guess

it was implied before but that basically confirms it

[elder rapids]

ye it's still up in the air

[keen beacon]

thats a generous interpretation of it imo

i hope so

[elder rapids]

it's gonna be hard to get a model that can select and identify those wording schemes and what the model is anticipating, the route it intends and it says it's going towards, the key words it's relying on, the aha moments to take advantage of

it's the little things and that's going to so hard for them

but that's accepting the premise of distillation, I disagree this is possible in AI studio, and summaries should only exist in api

[keen beacon]

yeah summarization inherently dilutes the signal

tbh it wont prevent sophisticated actors, openai seemingly has a lot of protections about it and even then its not flawless. google has even less

it makes the user experience more annoying generally

[elder rapids]

ye, there's really a lot of reasons summary is just a fleeting decision

rather than actually substantiated

[keen beacon]

you can still trivially leak the cot i guess, so ill be doing that if i need to understand exactly what the model is doing at times

its just that its potentially unneeded degradation

[narrow elbow]

This can be seen as a cheating behavior. The so-called thought summary and the answer itself may have no direct correlation (the model is not open source, and the correlation between the thought chain and the answer cannot be verified). Even if two models are used, one for thinking (and generating summaries) and one for answering, users cannot detect it. This makes the model lose reputation and trust (even if it is strong). Using security and preventing distillation as an excuse does not seem like something a super large enterprise would do. It is stingy.😏

[dusky aurora]

what is goldmane?

[elder rapids]

agi

[unborn ocean]

Let’s be real: we will all get used to it.

Big companies might get access, we might not.

[misty vault]

agi

[small haven]

omg its coming

[elder rapids]

btw 2.5 pro 0506 knows this too

[kind cloud]

Yes. So, strictly speaking, this isn't meant to find out 'goldmane'; rather, it's the way to exclude Flash models.

https://fixupx.com/AiBattle_/status/1926323159126487528

AiBattle (@AiBattle_)

Interesting behavior from the Gemini model Goldmane
︀︀
︀︀When prompting the Goldmane model to generate design concepts or plans, it often attempts to include multiple images in its response.
︀︀
︀︀This behavior is not present in the Redsword model, nor have I seen it in other Gemini models.

**💬 8 🔁 4 ❤️ 107 👁️ 8.0K **

▶ Play video

FixupX · May 24, 2025, 5:03 PM

Maybe the fact that it tries to cite images could also be a way to tell it apart, but I'm not certain yet.

[brittle hull]

oh

[kind cloud]

it's just a Chinese model

[brittle hull]

What makes u think that?

[kind cloud]

Because it revealed its developer to me

[small haven]

how do we do a poll here

[keen beacon]

plus icon -> create poll

[patent aspen]

IMO whichever one releases second will probably be stronger lol

[sonic tendon]

redsword leaderboard release wen eta

might not happen, but curious

[tall summit]

oh baidu

[small haven]

yes recency bias lol

[sonic tendon]

on that note, I've been wondering what folsom is

but yeah baidu and bytedance have been active lately

yeah

well, they tend to happen at the same time

[late path]

We will never see redsword again

[sonic tendon]

also, wasn't goldmane the worse one

[keen beacon]

redsword was removed

[small haven]

sure, i guess to some extent, it wont be dominating everything, but i do believe deepthink will edge in math/code a bit more, usamo at >40% is very impressive ngl

[keen beacon]

apparently not since it won over redsword i think

[sonic tendon]

"best" seems a tad subjective lol

they'll probably both have a case for being the best, I think

[small haven]

lmao

we have a strawberry in here

[sonic tendon]

what is a strawberry

I keep hearing that term

[keen beacon]

strawberry man?

idk

[sonic tendon]

the twitter guy

i assume

heh

[small haven]

my issue here is why would an actual google engineer be here to actually talk about insider info? what is there to be gain, other than attention and losing ur cushy job lol

[keen beacon]

nah hes just a massive google fan

[sonic tendon]

aren't we all

[small haven]

not wannabes

ok bud

[sonic tendon]

i doubt anyone cares that much

[floral merlin]

Hello, where is the price / score chart in the new UI?

This (Price Anaysis):

It is the most usable for me chart so far.

[keen beacon]

it might not be added yet in the new site idk

[floral merlin]

Ok, thanks for the info 👍

[echo aurora]

Yeah sry to say isn't currently on the regular site

[floral merlin]

No need to sorry. I know that an art of creating software is always an art of sacrifices.

[elder rapids]

goldmane was better than redsword in most of my tests

[keen beacon]

i havent used the model yet lmao i just read the chat 🤣

there were people saying both things i dont know

[jade egret]

hi

[elder rapids]

ion know bro, I've been accurate about basically every assessment ive projected here

which isn't much projections at all

but still

[echo aurora]

hello

[elder rapids]

I KNOW models

[misty vault]

do u know that gpt-4-0314 is agi

[elder rapids]

gpt 40 is agi

[golden ocean]

gpt 4 is agi

[high ginkgo]

then we go to a more recent model gpt 4o and we're back to narrow ai ☹️

[sturdy mica]

https://tenor.com/view/funny-cool-2023-funny-memes-hahaguffawchuckleleamalus-this-is-the-hahaguffawchuckleleamalus-gif-4886308197560020132

[small haven]

no comment

[sour spindle]

Opus 4 just took the top spot on SimpleBench

[primal orbit]

is opus in the lmarena direct chat thinking or not?

[torn mantle]

its good

although the radar chart is a bit broken visually

its def better than whatever they were providing previously

deep research feature or wtvr...

[elder rapids]

just about what I expected tbh

also I don't think opus 4 or sonnet 4 nonthinking are going to be much higher than 3.7 nonthinking

[misty vault]

real

[small haven]

these tweets about o3 pro is making me 🥵

@deep adder why can't i paste image into claude code anymore

it used to work

i guess they disabled it :/

or my settings is fcked

[hollow ocean]

@misty star Disappointment

[small haven]

claude code is a fcking beast

ya not ai related, if they can do, why can i post a xi pic

*cant

[small haven]

oh hell naw

wait i was using haiku?

tf

lmao, i was running multi agents

i think its caused i ran above limits

it defaults to haiku

i was wondering why i was getting shxtty results

[hollow ocean]

https://x.com/p4mui/status/1928234595385819605?s=46&t=AH7sIlIv16Z3Kdb6j3cjfg

P4mui  (@P4mui)

BREAKING: XChat is now rolling out in beta for a select few users

It will roll out to more users soon

As a reminder XChat is the new generation of DMs with new features including:

• End-to-end encryption

• Ability to send files

• Ability to unread!!!

• Ability to

X · May 29, 2025, 11:38 PM

[pseudo hemlock]

Is the "prompt to best model" feature gone?

[late path]

legacy.lmarena.ai

[pseudo hemlock]

I'm on there and went to "prompt-specific leaderboard", put in the prompt, and it doesn't load anything after I press send

I even tried https://github.com/lmarena/p2l and went to "Try on Chatbot Arena at the Prompt-to-Leaderboard tab!" and still nothing

[echo aurora]

like cap said isn't apart of the current site atm, I'm going to flag to the team regarding p2l on the legacy site. sorry for the inconvenience!

[pseudo hemlock]

its okay! thanks

and not a big deal, i'm sure i'm the only person who wants it enough to try it on the legacy site lmao

[elder rapids]

nah

it's actually pretty fun

[elder solar]

why i dont have this feature in my chatgpt logged account?

it shows off when logged out

[torn escarp]

Is the price analysis chart planned in the new UI?

[drifting thorn]

Flux Konnect is good as hell

[misty vault]

fluxsydney

[echo aurora]

That’s tbd, but good to know something you’d like to see brought to current site

[ocean vortex]

I don't think it is, but you can kinda sorta trick it by adding <thinking> at the end of your prompt

native thinking:

with a prompt, thinking 'disabled':

the thing with Opus is that it usually doesn't reason for long either way

write me a poem that doesn't rhyme. <thinking>

[echo aurora]

prompt-to-leaderboard is working again

[pliant cypress]

New deepseek R1 score 10% higher on SimpleBench wow

[torn mantle]

the new deepseek is actually crazy

its the closest model to o3 in terms of formatting

they are trying to mimic that

pretty sure its trained on o3 outputs too

[tall summit]

i hate o3 formatting

[torn mantle]

i love the arrows explanation, straight to the point

i myself use that a lot

[keen beacon]

nah this new r1 is closer to gemini

in terms of output

[torn mantle]

gemini is like the biggest yapper

it doesnt write with emojis nor arrows

but its packed with so much knowledge

so we got the formatting + knowledge from both models

[keen beacon]

https://eqbench.com/creative_writing.html (see new r1 and click the slop metric)

for creative writing probably personally

but eqbench's creative writing leaderboard is a useful metric

[torn mantle]

thats on creative writing

[dusky aurora]

I like yapping models

[tall summit]

↑

[keen beacon]

still the slop metric is somewhat useful to determine provenance

if u understand how its calculated

[dusky aurora]

opus writes so great scenes

[tall summit]

opus my beloved

[torn mantle]

lol

grok 3.5 is coming

just wait a year or two

[wintry tinsel]

It is lol

My metric is different though I jailbreak and unlock the model first, than I judge its full underlying ability

Opus is so far ahead its in a league of its own

Like a fundamentally different class of performance, I’ve struggled to step down even to sonnet after using it

For creative writing I mean

[small haven]

opus is built diff

[ocean vortex]

He finally added Opus as well 🧐

[ocean vortex]

I think Opus is easily the biggest available reasoning model right now tbh. That's not to say that it is the best overall since it clearly isn't, but there are things it's gonna take the lead at

[small haven]

omfg i have it

craig smart

7mins for news is crazy tho

i thought u was trolling but dayum, o1 pro in the api boys

*o3 pro

[tall summit]

i knowwwwwww

[wintry tinsel]

Lots of things I’d say its gotta take the lead on at least half of the things people do with LLM’s

[keen ferry]

you got pro subscription?

[torn mantle]

yes

[small haven]

no its a free account

[torn mantle]

stop it

or dont

idkdidkidkdi

[small haven]

its free, i just had to splurge $200

o3 pro is super long tho, way longer than the old o1 pro

[torn mantle]

what can we do without you?

first one that will have o3 pro access in this server

what a privilege

😖

[small haven]

ok lemme put ur berberine prompt in

[torn mantle]

lol

is it out yet or what?

[small haven]

not officially

[grim axle]

anyone having problems?

[elder rapids]

I hate this benchmark so much

😭

whoever made it has no idea how to judge + the model capabilities to judge

look at the bias control clarifications lmao

[small haven]

https://chatgpt.com/share/683b5183-2a90-8003-b84c-a73e47f0d345 @torn mantle

[elder rapids]

it is broken ye

[echo aurora]

Yeah I'm seeing this also, thank you!

Sorry for missing models everyone! Our team is looking into

[grim axle]

Claude 4 Opus was having problems too

[misty vault]

fr

[echo aurora]

Yeah I've been hearing the same as well, altho I can't repro because I can't access models 😭

[misty vault]

sorry for making the models unavailable

[small haven]

ok so wen is o4

and ultimately o4 pro

[grim axle]

it’s probably API fault

[misty vault]

no

if u haven't reloded since before the issue

then everything still works fine

[grim axle]

press ctrl w I found a way

[small haven]

oh my goodness, o3 pro is no joke

[civic flame]

@small haven you taking requests?

[small haven]

ok

[civic flame]

o3 surprises me by how meh it is at correctly formatting realistic wikipedia articles so try this:

"Write full Wikitext for a very realistic Wikipedia article for the 2028 Republican primaries, after the primary has finished."

[small haven]

queued

[split olive]

fix lmarena

[glad peak]

Anyone else having this issue? Was working fine 5 mins ago now all my chats are gone

[misty vault]

you got permanently banned

[split olive]

i get connection failed error

[glad peak]

Should've just scrolled up lol

[echo aurora]

Really sorry about that! Team is looking into a lot of widespread issues atm.

[glad peak]

Bruh you guys give me free access to all the best models with ease. It working 3% of the time would be a blessing

[small haven]

@civic flame what did you get usually for both candidates

their names

[tall summit]

well you have to show all your prompts to the world

[small haven]

https://chatgpt.com/share/683b6035-9634-8003-92e2-69522597c214 @civic flame

tim scott is that out of the scope

[grim axle]

Ask AI to fix the issues. Problem solved 👍

[echo aurora]

I shoulda thought of that! On it!

[grim axle]

It means you got banned

[echo aurora]

Lol not true!! ^

[split olive]

💀💀

[civic flame]

interesting

[grim axle]

What model should they release next?

[civic flame]

it went with a scenario i think is not particularly likely but all the same interesting to read

[misty vault]

gpt-4-0314

[civic flame]

unfortunately it looks like formatting is only a little better than o3's attempt

nonexistent template

[grim axle]

I don’t see why they didn’t added it yet

[civic flame]

skipped out a bunch 👎 lazy

[small haven]

damn

[misty vault]

openai not giving lmarena gpt-4-0314 access 😔

[small haven]

if 2028 is tim scott, im eating my shorts

[civic flame]

i think approximately zero of these people would decline to run lmao

claude 4 opus' prediction i think makes more sense, it went with vance

[small haven]

tbh vance seems like a sequential choice, pretty logical, but tim scott as the answer is diff and seems like it thought a bit for it

[civic flame]

i presume somewhere in the CoT it was like

"this is likely to be a very competitive primary without trump, and like in 2016 the winner tends to be hard to predict, so..."

which i suppose makes sense

shame they don't expose much of the CoT though

[echo aurora]

slight gestures towards

✅ Avoid political and religious content.

[misty vault]

I would vote for gpt-4-0314 if it were running for president ngl

[civic flame]

https://tenor.com/view/1984-gif-19260546

[small haven]

its ai related tho lol

[grim axle]

I need to try the new flux model please turn back on

also is the new model any good?

[elder rapids]

yo 2.5 pro in the app has better instruction following now

sum changed

[grim axle]

?

[keen beacon]

gemini product or aistudio? (or both)

[elder rapids]

it's interpreting my instructions and applying them in a much better way than before

[elder rapids]

product/the app

[keen beacon]

kk

[tall summit]

opus bad at translation compared to gemini 2.5 pro D:

[elder rapids]

oh ye they fixed the formatting issues on mobile too

[elder rapids]

sonnet imo is better for a lot of translation tasks

strangely enough

but 2.5 pro is a god at translation

same with 4o

[torn mantle]

nice one

[elder rapids]

although less nuance when pushed

[torn mantle]

unfortunately i cant tell if its any different from o1 pro

[elder rapids]

ask if it's agi

[torn mantle]

leo been playing roblox all day

[small haven]

o3 pro is less verbose

[elder rapids]

o3 is much less verbose in general tbh

[small haven]

its not a bad thing tbh

ya next year

lol

now we wait for deepthink :/

[elder rapids]

ngl high hopes for deepthink

if it actually pushes 2.5 pro further

[leaden palm]

you guys know you could code/orchestrate your own deepthink right

[elder rapids]

then that's something really to hope for

[small haven]

too lazy, i think im gonna put $250 in

[elder rapids]

bro acting like we're not the consumers 😭

[grim axle]

it’s pretty easy

[pseudo hemlock]

THATS WHY HES THE GOAT

[rigid crescent]

sorry if this has been asked to death already but is repochat planned for the new ui?

[echo aurora]

generally I won't be able to say if a specific feature is or isn't incoming; however, it's something we're putting thought into

[rigid crescent]

understood, thank you for clarifying 🙏

[small haven]

o3 pro uses big fat arrows 😮

[atomic pagoda]

Is anyone else getting the connection error?

[meager lintel]

New Gemini 2.5 pro checkpoint in a few days

Probably goldmane?

Where’d Tuesday come from?

😄

Well, I heard from someone else who saw it leaked from semi-public info too so makes sense

LMArena staff sandbagging the leaderboard update until Tuesday would be 💀

Just like Grok 3…

In this case, the other source I think is from something similar to the feature flags leak of Claude 4

[torn escarp]

100%.
I don't know anything else close to it that so distinctly compares price/performance of LLMs

[errant thorn]

it was working for a few secs just now but now its back down

[keen beacon]

they should call it o2 pro instead for maximum confusion

[small haven]

i literally have it 😭

o1 pro with search enabled kek

o3 pro gave me alpha omg

[patent aspen]

When is the last time someone working at OAI said they were still working on o3 pro?

Looks like April 16

[small haven]

yesterday

[patent aspen]

Oh where?

[small haven]

oai cpo

on x.com

[keen beacon]

https://xcancel.com/btibor91/status/1928522990599172141#m

[small haven]

theres no way ppl still think o3 pro is fake 😭

o3 pro vs baseline

ok so u saying o3 pro but integrated into gpt 5 lol

i mean its still o3 pro

just a router

[patent aspen]

It sounds like o3 pro is coming out eventually

[small haven]

i mean yes if u scroll up a bit

[patent aspen]

I know. I'm just reaffirming based on the posts above

[small haven]

99.99% confidence band lol

[echo aurora]

site should be up and working again btw 👍

[keen beacon]

gpt 5 release might coincide/be somewhat correlated with gpt 4.5 being shut down on the api too maybe

which is in july

[small haven]

day 1 with o3 pro

gpt5 at the end of the day is just a router

[keen beacon]

was it worth the wait lmao?

[small haven]

yes

[misty vault]

gpt-5-preview-0314

[small haven]

? i think its just going to explode more? bc majority is just using 4o as default

[misty vault]

true

[keen beacon]

they spent a lot of time on 4o (mid train). whilst 4.1 mini/etc are fresh. it seems 4o is gonna be used for a while

(they talked about this in a podcast btw about the mid-train/fresh train)

yeah

[small haven]

yea, no one cares about the rest lol

wb grok 3.5

even when elon ma has a black eye

i feel like grok 3.5 bigbrain is just going to at most match o3

u rlly believe that?

[keen beacon]

lmao

[patent aspen]

Is bigbrain some meme?

[small haven]

i can not vouch for this

[keen beacon]

its xai's version of o3 pro/deepthink

they named it bigbrain

[small haven]

oh right xai will release a $200/mo plan 🧠

lol ahh

[elder rapids]

4o is spitting out images embedded into the chat

man where is goldmane

cool that's when my big ass TV arrives

ion believe that tbh

[misty vault]

wtf is ion

[elder rapids]

I don't

[small haven]

are u saying gemini 2.5 pro updated version is gonna match goldmane

noice

[echo aurora]

hey after the site was down are you now seeing your chat history ( ) or are you NOT seeing your chat history ( )?

[elder rapids]

goldmane is an explanation god

it's crazy intelligent and it's subject to being influenced more now

less dogmatic when it comes to uncertain things at first and brute forces conclusions to be more certain

what it says

it's subject to being influenced more now

no I mean it's different from 0506

[elder rapids]

it thinks for a while and sometimes doesn't think at all

which in the cases It was in

was surprisingly appropriate

[patent aspen]

Is it less verbose than 0506? It's supposed to be

Or at least that was highly requested

[elder rapids]

much less

[patent aspen]

I'm excited. I haven't got around to trying it yet

[keen beacon]

I hope the not thinking bug is fixed

It's really annoying when it does it

[patent aspen]

What is the bug?

[misty vault]

I think the not thinking part is the bug like he literally just said

[keen beacon]

It just doesn't think before the reply

[small haven]

ok after playing with o3 pro for a few hours, its pure insanity

[keen beacon]

Long conversations

Since they exclude the previous thoughts iirc in prev turns, at some point the model just doesn't have the tendency to do it. Weirdly they don't prefill the thinking delimiter so it can just do that

Might be fixed since their logic will be different with the next update I think since they're adding the toggle

Yeah I don't think so

The 'fix' is to ask it to think it's just annoying to do so

Sometimes it won't work and you have to rephrase it in weird ways to get it to think etc

I looked at the latency for the first token so it's not visual

[grim axle]

Okay yeah the flux model sucks ass

[keen beacon]

It could be that but it mostly starts happening in long conversations, and I think the mechanism is as above. It's been a thing since flash thinking exp

Also it can think twice or get into thinking loops (multiple thinking blocks per reply) so I think it's a model thing

[keen beacon]

For me the thinking twice thing/etc happens sometimes when I ask it to think

#general message here's an instance of it doing it (fyi I was wrong here about it being a special token)

[tall summit]

REAL

[meager lintel]

whatever X-preview is sucks lol

[meager lintel]

thank god

after making 05/06 the ONLY option with no option to still use 03/25, I was half worried the thinking spam was some kind of intentional thing to pump out output tokens.....

yea makes sense

I'm hoping that whatever the "new research stuff" got into the new 2.5 flash is in goldmane too

new 2.5 flash is absolutely amazing

just not quite smart enough

but it hits so far above its weight it's insane

I saw some google researcher on twitter saying something like "a ton of new research ideas (which I can't talk about) were successful and got into 2.5 flash", so it's got my hopes up 😄

[elder rapids]

Logan confirmed it's coming in the next few weeks

[keen beacon]

He said that a few days earlier

They removed redsword I assume release is imminent

https://xcancel.com/OfficialLoganK/status/1927791900817317992#m

[patent aspen]

Weird that he said a couple weeks on May 28th

[elder rapids]

it's either you or him

ye

[keen beacon]

Maybe Logan is talking about an even later revision

But this update is substantial

It would be strange

[elder rapids]

I mean, in my eyes it's better than 0325 all around tbh

0506 still had the same capability but you had to prompt it more

goldmane simply just does it

[patent aspen]

tbh if Logan said a couple weeks on May 28th, I trust him more than myself

[elder rapids]

btw I believe 2.5 flash was an exception

they didn't actually serve it

as far as I know, I don't use vertex

[keen beacon]

Why remove redsword this early though

[elder rapids]

difference could've been major

[keen beacon]

Maybe but more time couldn't have hurt

[elder rapids]

hopefully

I anticipate googles releases a lot

[small haven]

against what model lol

[elder rapids]

goldmane vs 0506

[leaden palm]

idea is that where 0506 is worse than 0325, new gemini will be better

[grim axle]

FLUX AI IS HAVING PROBLEMS

[leaden palm]

curious why you're still on beta.lmarena.ai and why you're using an ai built for image to image generation as a textual ai

perhaps it's because it didn't force you to attach anything

[grim axle]

I use it to edit images like this

[leaden palm]

well you didn't attach an image there did you

probably why you're getting problems

it works fine for me with an image

[grim axle]

I tried and it’s still not working

[leaden palm]

then that's odd

[grim axle]

is there a limit?

because I used the model 20 times

[keen fulcrum]

Claude is making these graphs and Cursor isn't great at displaying them 😄

Flux kontext amazing
https://fixupx.com/AngryTomtweets/status/1928509452493246911

Angry Tom (@AngryTomtweets)

Today @bfl_ml dropped FLUX.1 Kontext, a new multimodal model that understands both image and text inputs.
︀︀
︀︀It's now available in @LTXStudio for you to try!
︀︀
︀︀Try here: ltx.studio

**💬 2 🔁 10 ❤️ 119 👁️ 10.6K **

▶ Play video

FixupX · May 30, 2025, 5:50 PM

[elder rapids]

lmao why does 2.5 flash thinking identify as Claude 4 sonnet, with all of the up to date Claude information on the arena

it says the model string too, like the regular Claude models

I've also been seeing different models act in a way that don't align with their personality

there's definitely a bug going on rn

where it's showing the wrong model name

or its routing to a different model

even the other models, like "Stephen" or "x-preview" are doing it

Claude is sometimes identifying as a Google model, too

yo this is DEFINITELY happening

[hollow ocean]

Is it better than opus 4

[elder rapids]

opus 4 isn't a very high standard so ye ofc

[small haven]

o3 pro currently has a 64k context window 😦

if deepthink matches o3 pro, but offers 1m context window, google wins

[hollow ocean]

will you switch to team google if deepthink does?

[calm sequoia]

Someone on twitter posted this interesting table

[small haven]

yes, im not a dickrider for any, i want to use the most frontier model

[late path]

they should add phantom and nebula too

the beginning of google's legendary comeback😁

[dusky aurora]

ah,so goldmane is a Gemini version?

[vernal meadow]

yes they both are and they are both very good

[dusky aurora]

so they say Goldmane will be relased i the coming days?

[alpine coral]

that seems v odd

i can't rememember the last time a google (or anthropic or oai for that matter) model identified itself as a model from a different lab.. and it happened repeatedly?

[primal orbit]

dragontail was better than claybrook, yet they picked claybrook.

[keen beacon]

If it identifies as Claude 4 sonnet at the very least the system prompts are all mixed up. Or the model names are switched/messed up or it's both

[alpine coral]

oh true...

that would fs be the most obvious / likely explanation

[keen beacon]

There were issues with lmarena earlier this is probably related

[sacred plaza]

anyone know what Ilya Sutskever has been up at his new startup?

[patent aspen]

Raising money

Using Google TPUs for research

tbh I don't expect any company starting so late to become relevant, although I also didn't expect DeepSeek or xAI, so take my word with a grain of salt

[misty vault]

me when gemini 2.5 pro

no gemini 2.5 pro is king

i learned from u

no

u worshipped gemini 2.5 pro in may still 🥰

[high ginkgo]

gemini 2.5 pro is cancer

[misty vault]

it is though

[sacred plaza]

Elaborate and you not going to wait for deep think?

[high ginkgo]

i am going to wait for deep think

[dusky aurora]

I still see nothing better in the Direct Chat list. Even the May version of Gemini is a viceroy compredto others

Opus is good but so laconic

somehow modern LLMs tend toward giving short, token-hoarding replies

[keen fulcrum]

Did anyone hear of https://sambanova.ai before?

[patent aspen]

You think SSI will be relevant?

[late path]

I heard their context window is too small, making them almost useless in practice

[late path]

I remember Ilya saying their first product would be safe ASI.
We won't see them until ASI

[patent aspen]

Wouldn't they just get outpaced by Google if that's their approach?

That's why I don't think it's particularly likely that new entrants will catch up

I think DeepSeek will remain relevant because it's based in China, and the US may eventually ban China from using US models

I don't think it will surpass the top US model in capability

[dusky aurora]

https://cloud.sambanova.ai/playground

[late path]

I'm not sure if every model Deepseek releases is only slightly worse than SOTA. Is this a coincidence, or is it because the distilled data they used fundamentally limits their upper bound?

[tall summit]

which is great

and it is wordy when you ask

[dusky aurora]

I prefer long replies. Curent Gemini is all about bullet points,which isreadable but too abrupt

[sacred plaza]

Ssi isn't trying to make money or sell stuff though. Why are you comparing these AI labs with SSI? It seems more like a research facility

[sacred plaza]

This seems probably false. Agree that they are competing on the same pool of resources when it comes to gpus though but the goal for SSI does not seem to be AGI

Very cool. Have not heard of blue sky research but would not be surprised given Google deepmind as institutional.

Given how hard it was to get even 20% of the compute for open AI for safety safety specific testing that led to the creation of anthropic, I don't see the market incentives promoting safety work for its own sake, like SSI.

Would be glad to be proved wrong tho!

Wish I knew that deep seek release before I brought the ultra plan, lol 😭. Agree that AI labs are doing safety work, which is definitely promising!

I agree with this take. It just seems like Google is adding graph of thoughts promoting technique into the internal model to create deepthink. This kind of seems similar to how AI Labs put in chain of thought prompting into their models to develop their reasoning models.

[feral lichen]

best ai for roblox studio?

[late path]

it would be a remarkable achievement if they could replicate the huge elo improvement that Alphago achieved by using mcts over raw DNN in LLM

[narrow elbow]

pursuit of technological dominance, tech giants and capitalists always prioritize capability over safety.this race mentality remains unchanged. just like cold war nukes.

[patent aspen]

They're terrified of safety incidents though

[narrow elbow]

yea

[elder rapids]

could be

but this is crazy tbh

[cedar tide]

when the next leaderboard update ? there are 6 models in the arena not yet in the leaderboard 🥴
(Two Claude 4, new R1, grok 3 mini,
qwen 3 no think, glm 4 air)

They added it to the battle arena recently

@deep adder
Now in battle arena

[leaden palm]

depends on how long a while is

[elder rapids]

couple days

[tulip meadow]

Hello can someone help me?

In Which section, Can I use as ai?

[cedar tide]

Nope

[torn mantle]

lmao

i thought the next grok model on lmarena will be the 3.5 ver

ig we just have to wait a little longer

[echo aurora]

at https://lmarena.ai/ you can use ai through: battle, side-by-side, and direct chat

[tulip meadow]

Thank you

Anyone have torrentleech site access? I need invitation

[verbal nimbus]

Would be cool if the Web Arena had a Svelte mode. Curious to see if the rankings would stay the same.

[keen fulcrum]

So no r2 for the foreseeable future=
How long will we be stuck on R1?

[small haven]

o3 pro before grok 3.5 is crazy

it happened

troll?

[elder rapids]

btw

I'm not getting goldmane NEARLY as much as in the legacy website

which is super strange tbh

[small haven]

@deep adder https://chatgpt.com/share/683c9650-6a04-8003-9b6f-d18411d02799

ok ya officially, who tf cares

its already here

yes

o3 pro + claude code is the meta

yo but imagine deepthink matches o3 pro, right.. but with 1m context window, that would go insane

[elder rapids]

ngl this WOULD go insane

[small haven]

no

[elder rapids]

but hold on I want to know

deepthink isn't going to be 2.5 pro 0506

with more thinking

it's going to be goldmane lvl probably

with parallel

[small haven]

its official name is literally gemini 2.5 pro + deep think

[elder rapids]

which could be crazier

[small haven]

exactly lmao

u can retire at any age

[elder rapids]

yeah?

[small haven]

thats why i keep myself busy

[surreal warren]

Any idea What tools to try for Deep research sites or scrapping
To Find me items matching specs

Chatgpt is fabricating

[small haven]

u can batch tasks in parallel in claude code, amazing

yup

u can run as many as u want, but for my case, 3 was enough

add coffee

[candid harbor]

try quadruple espressos

[feral lichen]

How can I continue a conversation if you keep standing like that?

[grim axle]

make a new chat

[hollow ocean]

GPT-5 July confirmed ✅

[elder rapids]

you can suspect it via the model deprecations

but ion think it's absolutely confirmed

[hollow ocean]

https://x.com/btibor91/status/1929241704873308253?s=46&t=AH7sIlIv16Z3Kdb6j3cjfg

Tibor Blaho (@btibor91)

@Angaisb_ July

X · Jun 1, 2025, 6:20 PM

[surreal creek]

is Stephen a different version of R1? it keeps answering in Chinese which is a pretty obvious giveaway, but there’s a May version of R1 not codenamed in the arena

unless it’s an undisclosed version of Qwen

[elder rapids]

it's not

they're just random small Chinese models

same as X preview

Gemini 2.7 soon

get it right

[patent aspen]

Knowing how models are named, it honestly wouldn't shock me

[elder rapids]

ion think Google would do that tbh

[patent aspen]

I don't either

[elder rapids]

they have some sort of philosophy of design

[drifting thorn]

Extremely good at image consistency

[drifting thorn]

How?

[leaden palm]

openai style:

const solve = async (prompt) => {
const results = await Promise.all(Array.from({length: 9}, () => generate(prompt)));
const index = parseInt(await generate(
  `We tried to figure out the answer to the prompt <prompt>${prompt}</prompt> 9 times. Write a final answer incorporating the best aspects from all of these: <answers>${results.join("\n\n---\n\n")}</answers>`
));
return results[index];
}

open source style:

const solve = async (prompt) => {
return generate(`${prompt}

Note: whenever you are about to end thinking, don't. Instead, first write out what you were about to respond with, then critique it in depth, then keep thinking. You are only allowed to end thinking once this has happened 5 times.`);
}

tree of thought style:

const solve = async (prompt, decisions) => {
const trials = (await generate(`You are in the process of solving the prompt <prompt>${prompt}</prompt>. You've made these decisions so far: <decisions>${decisions.join("\n\n---\n\n")}</decisions> You now need to either list some possible paths you can take (separated with the separator ---) or only list the final answer.`)).split("---").map(x => x.trim());
if (trials.length > 1) {
const results = await Promise.all(trials.map(d => solve(prompt, [...decisions, d])));
const best = await generate(`You are in the process of solving the prompt <prompt>${prompt}</prompt>. You've made the decisions <decisions>${decisions.join("\n\n---\n\n")}</decisions> so far. Now, you made some more decisions, resulting in these results: <results>${results.join("\n\n---\n\n")}</results> Write your final answer to this prompt, combining the best aspects of each.`);
return best;
}
return trials[0];
}

[wintry tinsel]

After opus my reaction to AI news has been

🫤 😕 😑 😐

Like they is nothing interesting happening

[leaden palm]

me after the sota isn't beaten for 11 days:

[hardy pecan]

we are in a golden age

but have trouble perceiving it

[leaden palm]

idk its said that logan is responsible for accelerating gemini's availability and ai studio dev

[surreal creek]

bahahahaha

[hollow ocean]

Opus sota

[small haven]

they alrdy nerfed o3 pro great

[elder rapids]

I mean it is kind of magical tbh

like, could you imagine being able to speak to something that isn't human but can actually, coherently, and with extraordinary articulation

talk to you about things that are extremely implicit, beyond the syntax it's built off of

and so reminiscent of human thoughts

and then now it can access and see your screen, and think about what it's looking at with the necessary context, to actual figure it out

rather than a narrow program that key logs then executes that in repetition

all in the span of a year

[keen fulcrum]

LLMs & Models:

DeepSeek: R1-0528 released: 64K context, efficient quantization.
OpenAI: Deprecating GPT-4 32k for GPT-4o, chat log & censorship concerns.
Anthropic: Claude Opus safety report, mechanistic interpretability tools.
Google: Veo 3 video model, SignGemma for sign language. Gemini 2.5 Pro: large context, UI/creative limits.
Mistral: Agents API for orchestration.
AI21: Jamba model reception good, details limited.
XAI: $300M for Grok on Telegram, skepticism remains.

Agents & Tools:

Perplexity: Labs for multi-tool workflows, new features.
LlamaIndex: Agents in Finance workshop, advanced RAG.
VerbalCodeAI: AI terminal tool for code analysis.
Latent Space: Collab on autonomous engineers.

Infra & Hardware:

Unsloth: Optimized DeepSeek models for limited hardware.
AMD: Max+ 365 GPU (128GB VRAM).
NVIDIA: Blackwell optimizations for DeepSeek R1.

Open Source:

Ollama: Naming issues, SDK instabilities.
Hugging Face: Diffusers enhanced, LightEval v0.10.

Challenges:

Cursor: Backlash over slow pool removal.
Manus: Instability, network issues.
Nomic.ai: Cloud security concerns.

New:

Black Forest Labs: New AI lab, Flux-1-Kontext image model.
Factory AI: Autonomous software engineers.

Insights:

Mary Meeker: AI industry report: accelerating adoption.
Microsoft: Early Sora API access.
Cohere: AI automation gains.
Gradio: MCP hackathon.

[misty vault]

gpt

[small haven]

I think they tried o4 pro

[willow grail]

ayo!!!!
TO ALL OUTSIDE VIBE CODERS!!!!

u know which laptops are best for vibe coding with our lovely cc/cm in parks?

[calm sequoia]

I think "vibe-coding" does not impact your laptop choice anyhow.

It all depends on what you code. Web - buy something nice to hold and use. Heavy workload (apps, data processing) - buy something powerful, even if ugly.

[willow grail]

so any website would be not heavy workload?

even amazon?

[calm sequoia]

Idk I havent seen website that requries a lot of cores or GPU

And amazon website is 💩 wdym?

[willow grail]

hmmm okey... but like lets go one step earlier..

how to vibe code with claude code on ANDROID?
do i even need a laptop to use claude code on ubuntu? or is screensharing from pc to outside device enough?

WE DONT KNOW

stuff like laptop and tablet is an issue cause it probably will be impossible to use without a table high enough

i am feeding crows at graveyard. this takes time. like 2 hours daily.

so i wanna code while at it

i could bring a mobile table with me if laptop/tablet

slept 8 hours R.

claude code is the best agentic vibe coder ou tthere

ok then why did my lower arms hurt a lot 10 years ago after 5h usage of macbook??

weighted 1.6kg

i can watch websites on android yes.

just websites and perhaps games up to phaser.js level

[misty vault]

no way

[willow grail]

like how? u would need to rotate your head a lot down

thats unnatural and thus will lead to pain

[patent aspen]

If I didn't have a full time job and had the energy and will to code in my free time, I'd probably switch from a Macbook Pro to a Framework running Arch Linux

[queen siren]

anyone recommend good libraries for running multiagent llm systems?

[patent aspen]

The Linux terminal feels like putting my hand on the third rail of the universe, and no amount of money, build quality, or brand value can give me that high

[glad jackal]

Guys currently which llm has the least hallucinations is it Gemini 2.5?

Lmarena?

What abt Gemini 2.5

It has Been 1 in leader board

[grim axle]

can’t wait until Claude opus is open source

[elder rapids]

it is Gemini 2.5, the reason why gpt 4.5 is said to not have much hallucinations is because it doesn't assert much

[elder rapids]

yes bro opus never hallucinates

Claude 4 opus is a hallucination goblin btw

[acoustic cliff]

[echo aurora]

going to be listening to lofi most of the day in #1340554757827461215 for anyone interested

[feral lichen]

can anyone tell me best ai for lua coding?

[warped sequoia]

you love lofi so much 😭

[balmy mist]

anything new?

[sour spindle]

Dumb question that’s probably been answered is the claude 4 opus on leaderboard thinking or base

[ocean vortex]

dork 4.0 agi confirmed

[echo aurora]

it is the base version

[sour spindle]

Pretty good score for the base version

[ocean vortex]

there's no such thing as claude base though, both are the same exact model

chat

just different prompt template 😉

And if you add <thinking> at the end of your prompt, from my experience this is gonna be largely the same as thinking natively enabled

[small haven]

same day as o3 pro official release 🥳

[ocean vortex]

I'm thinking they are to revise their implemention fundamentally maybe. Otherwise it makes no sense why it is taking them so long...

Straight forward parallel compute, you could implement that in a short evening lol

[keen beacon]

apparently o3 pro is already rolling out

[ocean vortex]

and it doesn't need safety testing additionally etc

[keen beacon]

just recently

[small haven]

i mean i have it alrdy, just need them to officially release, bc they kept tweaking it its kinda annoying

[small haven]

me

bro got amnesia

[ocean vortex]

I'm using o4-high early access

seems on pair with 4.0 dork

[small haven]

im not even trolling

[keen beacon]

hes not trolling

[ocean vortex]

are you not? Is it good then?

[small haven]

ok let me get some proof, sigh

gimme a prompt @ocean vortex

that involves the internet lol

[ocean vortex]

is it with tools or no tools?

[small haven]

no clue, it doesn't show explicitly in the cot

it does use web tho

[ocean vortex]

we need smth it can't cheat with tools on then... try this:

approximately:
A. 4km eastward
B. 30 km northward
C. >30km away north-westerly
D. <1 km northward
E. >30 km away north-easterly.
F. 5 km+ eastward
G. The glove is exactly where the car was at the time it slipped out.
H. Neither option is correct.```

[small haven]

queued

[marsh stratus]

is Claude 4 Opus on the leaderboard the thinking Claude or nonthinking?

[ocean vortex]

neither

seriously though read several messages above lol

[small haven]

pretty sure it does use tools in the backend, just doesnt show anything like o3

[ocean vortex]

it probably does. O3 non-pro early access did use tools as well

[patent aspen]

What is the source for o3 pro rolling out on Thursday?

[humble heath]

will o3 pro be added to the arena?? i assume not bc the model is so expensive

[small haven]

i have it alrdy stealth released 🤷‍♂️

and thursday is where big releases happen

[small haven]

https://chatgpt.com/share/683df8dd-3fa8-8003-a892-27d46a39e79c

[ocean vortex]

https://chatgpt.com/share/683dfaa6-ff4c-800b-bd59-1dc3642799a2 lmfao

[small haven]

it continues based off ur account, not mine

[ocean vortex]

yeah ik, but I don't have 3.5 obviously, it's still hilarious why it responded this way. It's just showing "chatgpt" for this chat no model name

[small haven]

let's be honest tho, do u really think 3.5 could have done that

[small haven]

is it even correct thats my q lol

[small haven]

no i meant the answer

is it A

[ocean vortex]

it's D

[small haven]

damnn

[ocean vortex]

I haven't seen any model get this right yet though. So it doesn't mean it's sht. Just that maybe it is not significantly better than normal o3

[small haven]

deepthink next

[small haven]

grok 3.5 on big brain could prolly crack it ngl

wen deepthink sir

yes

im actually curious

yes i believe that

its brian

[patent aspen]

It's a new person

[small haven]

nah but when is deepthink

gimme the exact date

[patent aspen]

Anat

I remember Ruth because she made our perks suck. I don't think about Anat much

She's kind of just there

[small haven]

brian gimme deepthink exact date release

[keen beacon]

he probably doesnt know xd

[small haven]

go into that deepmind office and asked everybody for the date

[keen beacon]

yeah but idk what you expect from him if u keep asking about it

[small haven]

hes talking quiet guys

so basically end of june

ur not in that slack group? come on bro

i have a genuine question, will deepthink have a 1m context window like the rest of the gemini models ?

[keen beacon]

theyre releasing two revs in a month?

or naming it ga later

i guess its the latter

[small haven]

wait they can actually afford deepthink on 2m context window? i believe it for regular gemini models

[keen beacon]

load i guess

[small haven]

im guessing keeping compute for research/training

we need an oai insider in here now

yoooo wen is o4 pro

ok bud

2m context window on deepthink is absolutely going to crush o3 pro, sorry sam

[keen beacon]

can deepthink use tools?

if not there will still be a use for o3 pro

[small haven]

maybe not, but 2m context window is a huge deal

o3 pro currently has 64k

i maxxed it out

128k got timed out

80k~ timed out

58k ish passed

original o1 pro could fit 128k

o1 as well

o3-mini-high too, but now it's limited to 64k

yes...

cuz of the huge spike in new members

google has no users, so 2m context window it is

no "loyal" users

😭

[patent aspen]

I mean that's their first mover advantage

It is

[small haven]

"forced"

nk style

[patent aspen]

You are correct Gemini will not be ChatGPT

[keen beacon]

4o image gen was probably bigger than 2.5 pro

[small haven]

400m in ai studio? or 400m android users who have no choice to pass by a gemini feature lol

geographic 100% india?

and other third world countries

i believe tides will turn when deepthink release all jokes aside

[patent aspen]

And you are a shining example @deep adder

[small haven]

craig singh

@patent aspen also when is jules actually going to be GA? would gladly pay for unlimited like codex

[patent aspen]

No clue

Your politeness is incomparable

[small haven]

its a fact, not conspiracy

western world is iphone dominated

[green oak]

I would like to but the buisness

I have 3 dollars

[small haven]

wtf

[patent aspen]

"third world country type" comes off a bit differently haha

Kind of like poors or plebs

It does. I'm just remarking on your word choice

Oh absolutely

More so in the United States among gen Z but yes

Sure

[small haven]

i think thats a you thing 😂

[ocean vortex]

agree

but the still thought $250 sub was great idea

[sour spindle]

Gemini app is atrocious

Normies love ChatGPT vast majority don’t even know there are other models

[misty vault]

https://tenor.com/view/cat-look-cat-look-at-camera-silly-cat-in-a-cage-gif-889392959852579879

[patent aspen]

People (especially rich people) have tended to value materialistic, branded status symbols less over time

Why do you think that?

Interesting. It does seem there was a counter-trend among millenials toward experiences, health, and well-being, although the current trend among younger people does seem to be towards materialism

[elder rapids]

this IS Gemini usage...

[red sluice]

Gemini pro is winning on most benchmarks

[hidden quartz]

Hey guys wanted to know if images generated from lmarena can be used commercially? And if not would editing it help??

[red sluice]

Depends on the models I guess but since the results are open source if someone really wants to prove it’s from their model, they can find and prove it’s theirs

[hidden quartz]

Umm currently using flux to generate product photos and then changing the subject with my own and keeping the background. Should I be worried?

[zinc ore]

AI Overviews is 1.5b if they were counting that. Per IO

I think Meta claims their AI usage is 1b (or close to that figure)