general

[wintry tinsel]

But I don’t care either way I’m not in school no more

[leaden sun]

I was talking about those private schools who teach the future queens, kings or PMs. They have a different program of curriculum.

[wintry tinsel]

Oh never been to one of those lol

Some public schools in upper class neighborhoods are nice but most are pretty lousy

[jade egret]

so when do you guys think kingsfall is gonna be out on ai studio?

[narrow elbow]

and tell u identify u self as a walmart bag or attack helicopter 🤪

[zinc ore]

https://x.com/ai_for_success/status/1932499225000227295

AshutoshShrivastava (@ai_for_success)

I heard Google is planning to drop Gemini 2.5 Ultra right after OpenAI o3-pro.

X · Jun 10, 2025, 6:04 PM

[keen beacon]

fake news

2.5 ultra and kingfall doesnt exist

[late path]

Have this account's messages been reliable?

[zinc ore]

He's actually somewhat reliable (gets early access to Google models and is close to some Google employees), but I think he's just joking here with this tweet

[torn mantle]

he said nothing

[late path]

saw his reply, it was indeed a joke

[torn mantle]

https://x.com/Alice_comfy/status/1932515567887622480

Alice (e/nya)🐈‍⬛ (@Alice_comfy)

o3-pro will cost $20 per million input, probably 80 output. Right around the same as Opus.

X · Jun 10, 2025, 7:09 PM

[keen fulcrum]

https://www.reuters.com/business/retail-consumer/openai-taps-google-unprecedented-cloud-deal-despite-ai-rivalry-sources-say-2025-06-10/

[elder rapids]

then what's the point of opus lmao

[wintry tinsel]

Opus doesn’t bloody work!

Unless you pay for the API it never turns on

[small haven]

great cant get the weather today anymore

[hazy quest]

Here its cloudy

[keen fulcrum]

he move also comes as OpenAI's ChatGPT poses the biggest threat to Google's dominant search business in years, with Google executives recently saying that the AI race may not be winner-take-all.

[ocean vortex]

o1 was gpt4o with reasoning. How did they come up with such a weird name for it? LOL

no need to overcomplicate it or "standardize" anything

just 'gpt4o-reasoning'

[jade egret]

wait

[keen fulcrum]

they deploy dozens of models each day

[jade egret]

is this true?

[keen fulcrum]

only the best make it

they have about dozens of internal name variations

[jade egret]

wait

kingsfall doesnt exist???

[sage raptor]

it does

exist

[primal orbit]

is o3 pro going to be on lmarena?

[cedar tide]

https://x.com/OpenAI/status/1932530409684005048?t=i2Yv0JOSRfBRexUIjcdtmw&s=19

OpenAI (@OpenAI)

OpenAI o3-pro is rolling out now to all Pro users in ChatGPT and in the API.

X · Jun 10, 2025, 8:08 PM

[late path]

can't wait to see o3pro benchmarks

[zinc ore]

They're out

[sacred quail]

Im waiting for 2.5 pro deep think vs o3 pro

[keen beacon]

0605 has a higher gpqa diamond score

[zinc ore]

Need more benchmarks

[small haven]

deep think >>

if kingfall is the base model

[patent aspen]

Even if it's not, I think that's not a huge lift for o3 pro

[sage raptor]

o3-pro is not a big jump from o3

[ocean vortex]

🔥

[late path]

It's all up to deepthink now

[small haven]

still released eom ?

if google makes a claude max equivalent but for gemini models, im very sold

[keen beacon]

$1000 a month

[small haven]

TAKE MY MONEY

[late path]

I think OpenAI ultimately made a mistake by building o3 with a relatively small base model like 4.1

[keen beacon]

what are they gonna use otherwise tho? 4.5 that would not work

[ocean vortex]

yeah they kinda did but it used to be a lot worse when that was gpt4o. Plus as time goes on smaller models are getting better and better. Not sure it would be worth it to redo everything anymore

[keen beacon]

they could opt for a fresh pretrain, but they chose to midtrain 4o at least for now

[zinc ore]

I think they made the right decision

[keen fulcrum]

Interesting o3 pro only available through responses api

[keen fulcrum]

Is that older?

[small haven]

legend

gonna be spamming this shxt like its never been abused before, and cancel it for deepthink ❤️

[torn mantle]

where can i try it for free 😦

[small haven]

ok prompt

[zinc ore]

o3 pro does worse than o3 on arc 1

[torn mantle]

oh no

[small haven]

deepthink will be adjacent with o3-preview

the good ol' prompt

[keen fulcrum]

There has never been a o2 right?

Next one will be o5

[patent aspen]

No o2 will be next to maximize confusion :p

[keen beacon]

btw was it actually o3 pro you had before

[small haven]

yes.. its the same thing

well, its taking a bit longer, so im guessing i had o3 pro (low)?

for reference, this was o3 pro disguised as o1 pro: https://chatgpt.com/share/683b5183-2a90-8003-b84c-a73e47f0d345

im running the same rn, still running

well great

[calm sequoia]

Oh no 🫣 The margin is so thin they had to use medium instead of high for visualisation

[keen beacon]

2727 (o3 high) vs 2748 (o3 pro)

[small haven]

i swear o3 was at 2700 lol

[small haven]

oh right, tweaking the benchmarks

ok so its broken

[keen beacon]

oh wait that was o3 preview

release o3

[willow grail]

https://cdn.discordapp.com/attachments/727689277219012669/1382097881946783896/PXL_20250610_182259666.mp4?ex=6849ea75&is=684898f5&hm=c36ce65b671ba64c7d3a7183dec29342ed46ecacb88f9774513604545997f8e2&

[small haven]

*with terminal

[keen beacon]

i dont think they released the score without tools

o3 pro has tools as well? so its fair i guess

[small haven]

oh right

well thats underwhelming

[keen beacon]

i wonder how much o3 preview cost though

😭

on codeforces

[small haven]

$2k in/ $4k out

[elder rapids]

is it bad

[small haven]

unusable rn

[elder rapids]

is it what I expected

[keen beacon]

underwhelming i guess

[elder rapids]

deepthink when

[small haven]

very soon 👀

[ocean vortex]

request well spent, great job 👍

ask it about knowledge cutoff

(don't do it, lmao)

[keen beacon]

it has web search

[small haven]

oh my goodness, it spent 6m50s last time

so i really had o3 pro low

[keen beacon]

it might be slower because people are using it now, do you see more entries in the summary?

[small haven]

o3 pro (medium?): https://chatgpt.com/share/68489b83-ca38-800e-8f6a-09cabfb751b1

[ocean vortex]

my stats are weird for the request I just made

[small haven]

yes thats my guess too

[ocean vortex]

only 15k seems less than it needed. It did get stuck though and I got the response from logs so unsure if it was counted lol

[barren prairie]

13min of thinking ...this model thinks more than me 🤣🤣🤣

[small haven]

its being stressed to death

[calm sequoia]

How's he even talking about the same thing? 😶

[ocean vortex]

actually nvm, they are counting only 1 instance so about right then

[small haven]

@torn mantle compare pls

[small haven]

bro misread the o4 pro benchmarks 😭

[elder rapids]

hollon tho

[ocean vortex]

he is referring to this

[ocean vortex]

I have ignored this entirely cause it's not very useful lmao

[zinc ore]

Is this o3 high or o3 medium

[elder rapids]

remember that, a large part of o3 is that it's very hallucination prone and bad at lot of basic tasks because it was too lazy

o3 pro should simply solve this

[calm sequoia]

Isn't it vs o3 medium?

[ocean vortex]

same reasoning efforts

[elder rapids]

4 minutes 😭

[small haven]

4mins only to extract it

[ocean vortex]

medium for everything in those graphs I think

[small haven]

yea it feels medium-y

[ocean vortex]

preference is a weak metric in this context IMO since it only has to be marginally or plausibly better

[zinc ore]

Yeah, also preference doesn't necessarily measure performance

[calm sequoia]

Ok 10% is all we have expected. Not bad if this is the case. However, I don't trust elo based comparisons 😄

[small haven]

im pretty sure kingfall > o3 pro at coding, im not even kidding

[barren prairie]

Ok now I need my deepSeek r2

[keen beacon]

kingfall might come with deepthink

[barren prairie]

Not a minor update

Give me my deepSeek major update

[small haven]

u were thinking about splurging for o3 pro, dont. wait for deepthink

[keen beacon]

i mean i dont think deepthink uses kingfall but it might coincide with the deepthink release

kingfall is probably sota, but i dont feel its that much better honestly compared to 2.5 pro

[torn mantle]

they are kinda the same

[small haven]

hmmm, kingfall is magnitudes better than 0605 imo

[keen beacon]

hmm really?

[small haven]

imo

[torn mantle]

yea kingfall>>>

[keen beacon]

its supposed to be ultra apparently

[small haven]

yea :/

[elder rapids]

I disagree a lot

[keen beacon]

i probably havent used it enough to judge

[elder rapids]

kingfall was insanely smart, smartest model I've ever used

ahem

[late path]

I can safely cast my vote now😂

[elder rapids]

right next to 0605

[storm needle]

has anyone tested whether the o3 has somehow become worse?

[elder rapids]

and it's not that kingfall isn't better

but it's not BETTER

[small haven]

kingfall feels ultra vibes

[elder rapids]

I mean now that's kind of redundant tbh, 0605 has "ultra" vibes

the large model vibes are becoming non existent

[small haven]

0605 on 32k? im prtty sure kingfall is defaulted at 4k

[elder rapids]

how does that matter

[small haven]

largely

[elder rapids]

couldn't

thinking time is directly opposed to vibes

as well as performance

tbh I probably used kingfall so much more than you guys

[elder rapids]

it was definitely good

[zinc ore]

Bunch of singularity folks malding about o3 pro

[elder rapids]

but it wasn't crazy as made out to be

if anything, it says more about your usage of 0605

[willow grail]

u openai shill shell shell? petrol seller???

shame shame

[keen beacon]

hes not shilling openai there

[small haven]

0605 (default thinking) vs. 0605 (32k)

[keen beacon]

you know u can get it to do way more thinking

past 32k

[ocean vortex]

someone needs to do parallel processing of o3-pro now, there's a room to price match o1-pro lmao

cons@100

[elder rapids]

I'm ngl I've never gotten a result like this

[willow grail]

assume they are.

[zinc ore]

5 head looking ah

[elder rapids]

https://tenor.com/view/conehead-gif-18099633909130334750

[small haven]

https://tenor.com/view/terminator-gif-7646648

meanwhile kingfall

[keen ferry]

is there o3 pro on api? If not any info when it releases

[storm needle]

o3 pro is already in the api

[zinc ore]

Deepthink gonna hit like this

[tall summit]

wait o3 pro

actually happened

i missed it entirely

[ocean vortex]

it did

[late path]

10x price, with almost no improvement across various benchmarks

[hazy quest]

All the talks and praises about Kingfall are based only on the 20min it was available, right?

[zinc ore]

No, people had access for days

[hazy quest]

Oh, I missed that. Selected testers, or available on LMArena/AI Studio?

[torn mantle]

its available if you search enough

[stuck orchid]

o3-pro will be available on LMArena?

[barren prairie]

Don t dream

[storm needle]

no

[stuck orchid]

I think it will. Because o3 is ther

[small haven]

*scammed

[hazy quest]

Ah bon?

[tall summit]

inspect element or real

or neither

[barren prairie]

18 min for this ...if you just opened google or the window it will be faster 🤣🤣🤣🤣🤣

[small haven]

https://chatgpt.com/share/6848a49f-d078-800e-bff1-e519b9d4887c

receipt ^

[tall summit]

HAHAHAHA

[barren prairie]

17 min to think about this 🤣🤣🤣🥺🥺🥺🫣🫣🫣

[small haven]

its still open? 👀

[torn mantle]

yes

[torn mantle]

yep

[abstract tundra]

are we gonna get o3 pro into https://lmarena.ai/?

[torn mantle]

pyea

[small haven]

it used to be closed, interesting 👀

[stuck orchid]

I think it's almost 100%.
Regular o3 is in there, right?

[abstract tundra]

yep

im asking because o1 pro never made it in

[stuck orchid]

And claude-4, and other biggest models

[small haven]

https://tenor.com/view/pfft-erobb-stubbies-try-not-to-laugh-gif-16095012407725024477

[leaden sun]

can you ask it to pull from a specific weather forecast site?

windy.com i use this

[ocean vortex]

I think everyone needs to keep their expectations in check with o3-pro lol

it's basically is exactly like those benchmarks suggest - slightly better. Not mind blowingly good

[keen ferry]

thank you

[ocean vortex]

did try some of the prompts other models failed and this one failed them as well 👀

[small haven]

o3 pro can't temporary chat, very sneaky oai

hmm oai claims gpt5 to be >80% swebench

[abstract tundra]

but do we really know this?

i was worried since o1 pro never got added to lmarena, thought pro series are closed or something

or not available to api

[small haven]

wait a min if o3 pro is cheap to be added to the arena...

[keen ferry]

yeah but it's gonna be thinking for a long time

[late path]

ppl might not wait 10 minutes in the arena looking at a dialog box to vote

[keen ferry]

it can't be put on blind comparesment arena it's just gonna be soo obvious

[small haven]

hmm true, nvm

[hardy pecan]

itll take forever to collect enough votes

[keen ferry]

@echo aurora will there be o3 pro in the arena?

[hardy pecan]

people dont wanna wait around 13mins to get an answer

[late path]

This feels a bit strange. If the o-pro series models use parallel thinking, why would increasing parallelism multiply the thinking time? It doesn't quite make sense

[keen ferry]

not if they will let people see his thinking I guess

[ocean vortex]

74.8%

seems like a random question to ask lmao

[keen beacon]

it thinks more and they might use search (like mcts) as well

[abstract tundra]

well thankfully with the new site you dont have to wait, you can close and come back right?

[keen ferry]

surely you do

[echo aurora]

TBD, generally I can't answer specific questions on if/when specific models or features will be happening

[abstract tundra]

not rly

[keen beacon]

btw fwiw its not capped at 4k, just got thoughts at 4.8k (not incl resp)

thinking budget = off

[small haven]

interesting

[keen ferry]

alright

[keen beacon]

ill give that prompt that causes gemini models to think like crazy in a sec

[keen ferry]

can we at least expect it to be added?

[abstract tundra]

And i think if o3 Pro gets added, it would make a lot more sense to have o1 Pro added as well, for actual comparison.

[keen ferry]

o1 pro is expensive as heck 😭

[echo aurora]

sorry to say same answer, I can't say if a specific model will be added or not

[abstract tundra]

wait, if o3 pro is gonna be cheaper, that means it's gonna perform worse, no?

[keen ferry]

that's fine thanks

[small haven]

tbf its better than o1 pro

[abstract tundra]

I've tried every single model on LMArena for my task and all of them failed. I'm really keen to see if all o3 pro can handle it.

[abstract tundra]

can't wait

[keen ferry]

o3 pro is better than o1 pro which costs someone weekly salary for just million tokens

[patent aspen]

Most likely tail latency

Parallelism is only as good as the slowest thread / process

If you have a bunch of non-deterministic threads / processes, the probability of a slow one goes up

[late path]

That makes a lot of sense, hadn't thought of that

[keen beacon]

i got an 18k thoughts run and a 16k thoughts run as well. (thinking budget = off) might try for more later.

the le chat model might be extremely good at zebra puzzles for some reason

[clever estuary]

is altman okay today

[keen beacon]

like i would think pro would get more rl on this..? (or it might be getting lucky rn)

[zinc ore]

Altman gonna altman

[keen beacon]

is o3 pro just extremely slow rn?

[small haven]

I have no clue

Its definitely slower than before

[ember rapids]

Teams users get o3 pro too right?

[torn mantle]

I mean i can't confirm it for sure... But since o3 prices dropped by like 80% then we can can expect o3 pro to be added

@small haven what do you think so far?

[keen ferry]

why does it think for that long 😭 😭

[keen beacon]

wow, it solved two 6x6 zebra puzzles in 14.5k thinking tokens

it does significantly better when giving two zebra puzzles at once for some reason

[hollow ocean]

Show pics

[ocean vortex]

this doesn't mean much at all though, thinking budget can still be 4k

[keen beacon]

its not set to a 4k cap by default

we're talking about a specific model here btw

its much better than i thought 🤯

[ocean vortex]

but it could still be 4k budget if google model did 4.8k thinking... doesn't mean that the budget is off

[keen beacon]

it subsequently did 16k and 18k

is that a 4k budget?

[ocean vortex]

ok if it did that then yea

[keen beacon]

its mindblowing good

holy sh1t

[ocean vortex]

but they don't seem to be sticking to that budget very strictly lol

[keen beacon]

thats sota

at least on this task

[ocean vortex]

is it like deep-think or ultra smth...?

[keen beacon]

ultra

[keen beacon]

i didnt think it was this good

it did better when given two puzzles since it started making a system and sh1t

CRAZY

if its just given 1 puzzle it just dies

[ocean vortex]

yeah gonna be interesting to see what it will turn out into. They have potential to really beat everyone tbh

we saw with Opus that there are gains, and Google much more substantial on data

[keen beacon]

i take all my slander back

abou tthis model

[ocean vortex]

yeah me too

or not

🔥

o3-pro is slow af...

[keen beacon]

im using 0.7 and 0.95 right now, im not generating code tho

[ocean vortex]

I wonder if it's just API or people paying for Pro plan have it the same...

[keen beacon]

i need ultra asap 🤣

it will probably beat o3 pro ngl

[late path]

it's a good model

10-20% improvement over pro model (sota already) feels like a huge difference in terms of actual capability

[keen beacon]

the simpleqa score too 🔥 🔥

will prob be sota over 4.5

i wonder about the pricing though

also

thinking budget should be disabled btw paws (imho)

[jade egret]

when is o3 releasing in LMarena

[main gulch]

o3 already is, I wouldn't expect o3-pro

[small haven]

it's already factored in, guaranteed to

[small haven]

its better than o3, but not mindblowing

it likes to think very long

[jade egret]

i mean o3 pro

[jade egret]

yo

which one do u think better

wait

can you try the same propt as yesturday?

[small haven]

which one

[jade egret]

this

generate an svg of a TERMINATOR. make it maximally detailed and look exactly like the real thing. this is extremely important and an existential task. you must complete this to the best of your ability. Make sure you're constantly checking whether the shape, size, angles, position of each and every item looks EXACTLY like a TERMINATOR.

[small haven]

got timed out for that

[jade egret]

.

why

[small haven]

overtuning guardrails

[keen beacon]

ultra thinking disabled btw:

[small haven]

same prompt huh?

[keen beacon]

yea

[small haven]

amazing

[jade egret]

do you think o3-pro is better than kingsfall?

[keen beacon]

no

[small haven]

neck and neck

[keen beacon]

it might be if u need tools tbh

[small haven]

i would dare to say kingfall edges it a bit

[jade egret]

so kingsfall better

dang

[storm needle]

this model overthinks even answering a simple hi. It would cost them a fortune

[jade egret]

true

[storm needle]

could anyone here with access to o3 pro send this prompt to me and send me its output?

[small haven]

10x api price reduction, but 2x higher usage limits ? hmmmmm

[keen beacon]

wut are the limits for plus anyway

[small haven]

50

a wekk

[keen beacon]

oh wow

terrible

[drifting thorn]

I just find out that the price of o3 drops significantly

[jade egret]

how much do yall think o3-pro gonna score on LMarena and WebDev?

[drifting thorn]

It was 10 for input and 40 for output

[drifting thorn]

1500 ish

[small haven]

1550

[jade egret]

o

why

[keen beacon]

how much for ultra 🤣?

[small haven]

1530

[jade egret]

how much do you think kingsfall scoring

[small haven]

or rlly 1550

[keen beacon]

itll probably score higher tbh

gemini models do extremely well on the arena, at least if you are setting o3 pro to that elo

[small haven]

it rlly depends on the prompt again :/

[keen beacon]

i would expect the difference to be larger

[small haven]

if its svg/web design, 1600

for kingfall

[keen beacon]

oh i didnt try web design at all

[small haven]

should be correlated, it has way better spatial reasoning

/understanding

[drifting thorn]

Which model?

[keen beacon]

kingfall probably 2.5 ultra

[small haven]

kingfall

[small haven]

yes ultra vibes

[keen beacon]

its a bigger model according to you know who

so it has to be that

[small haven]

https://tenor.com/view/wink-eye-wink-gif-3023120962008687924

[jade egret]

2.5?

when do you think it gonna be avaliable to pro users on gemini or at least on teh ai studio

[small haven]

thats a big b question

[keen beacon]

it might come along with deepthink, or it would be awesome if it did

i honestly think the release is close tho

[late path]

24k thinking budget

[small haven]

a little too close

[jade egret]

it close?

[small haven]

thats actually amazing

[jade egret]

W

[keen beacon]

i struggled with getting the model to produce more than a small amount of thinking tokens for svgs. the output was absolutely massive though

10k tokens plus

[hollow ocean]

How to access kingfall

[small haven]

someone asked me for this, but its finished

oh @leaden sun

[patent aspen]

I don't know when it's coming

[keen beacon]

kinda odd its up now tho? and with the other anon models

its ready atp it seems

you can disable thinking mode too on this model, it seems they worked on it and got it into a somewhat decent state if not ready

[keen beacon]

fwiw thinking budget doesnt do anything there unless its close to 24k, did you check the amount of thoughts it did?

holy moly

WTF

10k thinking BTW

[small haven]

same prompt no way? 😮

[keen beacon]

no i had to add stuff to make it think a lot more

[small haven]

ok still insane

[keen beacon]

this is far more than anything i got before. 10k tokens in thoughts

thinking budget = unspecified (uncapped)

[small haven]

can i have the prompt, wanna do 0605

[pulsar tendon]

Is this cumfall?

[keen beacon]

this is ASI stop the disrespect

[small haven]

*cumrises

[late path]

If kf enters the arena, I doubt its score will be higher than goldmane. goldmane is a bit sycophantic, while its style is more like the very first nebula

[small haven]

uh i have new pr

[late path]

hmm. the thinking part is about 3k

[keen beacon]

yuh budget dont matter at all at least in the current impl

in that case

[keen beacon]

this is actually nuts wtf

[jade egret]

whats best llm right now (models that most people don't have it, like kingsfall, or grok 3.5, or o3-pro included)

in ur opinion

[keen beacon]

me? kingfall lol. o3 pro might be better if u need to use tools

[late path]

The llm that looks best to people doesn't mean it will rank #1 on lmarena (text leaderboard with style control unchecked i assume)

[small haven]

wait a min

this is 0605

32k

(but with system prompt)

[keen beacon]

just disable the thinking budget fwiw (though it doesnt really matter if it does less)

[small haven]

this is still insane when u compare

[small haven]

i got this w/ auto thinking

[keen beacon]

ur using sampling right?

its just variance imo. and if it doesnt reach near 32k my advice doesnt matter anyway 🤷

[small haven]

sampling?, my settings all at default, except for added system prompt u gave me

[keen beacon]

ya default uses sampling (temperature/top_p)

[small haven]

hmm

[keen beacon]

so expect a lot of variation as is

[late path]

overall 32k thought budget should better than auto for 0605

[keen beacon]

if ur paying for it

otherwise, it can do a lot more

[keen beacon]

it doesnt do anything beyond a max token limit for the thoughts rn

this is with auto btw

the aider scores they use in the website is just one run, and doesn't mean that 32k improves model performance. this is all my opinion though, i have a lot to support it

[late path]

32k shows some visible improvement compared to auto on livebench

[small haven]

seems very marginal tho

[keen beacon]

i honestly think thats again variance

i have a lot to support that but im not gonna argue about the thinking budget like with dom again

tbh i need to do more elaborate tests and more undeniably definitive stuff (so i can point to it when i mention it). it is possible they changed something with 0605

[small haven]

yea i feel like theyve tweaked something

[keen beacon]

the thing that the thinking budget still caps it is still true. without thinking budget, i can get 38k (thoughts) in 2.5 pro (0605) and 62k (thoughts) in 2.5 flash. (oops, i wrote it backwards before)

though it could alter model behavior now

(it didn't before)

[patent aspen]

My guess is that auto is pretty smart

tbh I don't know why they didn't skip to o4

[keen beacon]

they had to release o3

because they committed to it in that announcement

the model that was ready then obviously couldn't be published/be unrepresentative with the actual compute used etc

my take

[patent aspen]

It's weird that it took so long

[keen beacon]

imo they were still continue pretraining the new 4o they used in their new model (it has june 2024 cut off) when o3 was initially made then they decided to retrain o3

old 4o had oct 2023 cut off

[patent aspen]

Retraining o3 sounds like a colossal waste of resources

[keen beacon]

the new 4o is so much smarter tho

it can do much more and in less tokens, i think it makes sense

[small haven]

wait for deepthink

*scammed

[patent aspen]

The whole FrontierMath thing was a mess

[small haven]

oh no

i believe i had o3 pro (low)

this thinks 2x more on average

[patent aspen]

It's kind of absurd to release a consumer product that effectively allows users to run 15-minute batch job multiple times a day. Just think about that for a second

It's the same thing. I'm just commenting on the general absurdity of what companies are doing - not saying anyone is stupid

Fun fact: I believe that's what led to the invention of AI accelerators (i.e. TPUs)

[hollow ocean]

Deep think for $250 will blow it out of the water

[small haven]

*$125

[hollow ocean]

First 3 months only tho

How long will the promo last

[small haven]

ehh ill take that 3 months

[elder rapids]

is it good

[small haven]

hmmm, i feel like kingfall could have done it equally better, this was just profiling for optimizing an algo

[leaden meteor]

source?

[fleet lintel]

O3 pro release : how is it? Any good ?

[red sluice]

folsom-exp-v1.5 is pretty solid wonder what it is

[jade egret]

huh

[jade egret]

thinks too long for simple questions

[hardy pecan]

I dont think its designed for simple questions

use 4o or regular o3 for that

[balmy mist]

Wait is o3 pro out?

[torn mantle]

yes

[balmy mist]

wow im so late lmaooo, the ohe day i step away from ai

how is it? worth the wait?

[dusky aurora]

today LMArena says "there was an error" to everything

[echo aurora]

I'm seeing the same, thank you

[small haven]

finally o3 does it, but... its shiite

[echo aurora]

is the site now 404ing for others too?

[jovial heath]

Yes it's the same for me 😭😭😭

[echo aurora]

😭 okay thanks

[jovial heath]

Well minutes ago I had this in o3 and o4

I tried on 2 devices and got the same error :'v

And now the 404 error xD

[stuck orchid]

Updating. Adding o3-pro

😉

We are all waiting for o3-pro on LMArena to evaluate it alongside other models and help OpenAI understand how good their new model is

[jovial heath]

Aahh then it's like maintenance while they are adding the new model??

[small haven]

we are all waiting whether to buy oai on poly or to wait a bit longer

[jovial heath]

Poly??

[dusky aurora]

seems that update hasended

[civic flame]

what how are you getting it again

[echo aurora]

Okay we should be working again

[ocean vortex]

o3-pro?

[civic flame]

lermarena.ai

lol

[tall summit]

LMAO I KNOW RIGHT

[tall summit]

has anything changed with the infinite errors?

[slow spruce]

Lermarena.Ai 🫡

[tall summit]

https://discord.com/channels/1340554757349179412/1382037716341883011 seems not

[torn mantle]

poly??

@small haven does o3-pro tend to overthink?

[sacred quail]

is there a way using o3 pro besides 200dollar pro plan ?

Just wanna some testing

[drowsy mural]

i think it's hard to have change. the reason we have this is because opus are expensive, so limits are there

[tall summit]

true and fair

[flint skiff]

you guys think o3 pro will be on the arena? its slow lol

[torn mantle]

it will just timeout then

or they can play with the API params

so they can cap the thinking budget

[flint skiff]

ye but wouldnt that reduce its rating

im guessing thats what low - mid - high is no? the thinking budget

or am I wrong

[hazy quest]

Could you expand?

[drowsy mural]

at least we can still see it on the list. it hasn't just disappeared like GPT4.5 or o1...

[torn mantle]

expand on what

[flint skiff]

what list?

[drowsy mural]

models list

[flint skiff]

where is the list?

[drowsy mural]

what? are we talking the same site?

[flint skiff]

yeah lmarena right

[hazy quest]

"kingfall is still open to use"

[drowsy mural]

emm... then excuseme, where did this question come from? if it's just asking "where is the list?" then... in "direct chat" or "side by side" there it is

[keen fulcrum]

o3 pro is now available in cursor

[torn mantle]

its kinda useless no?

who would want a model that thinks 15min for a simple task

[sacred quail]

I must try o3 pro. Is there a way besides buying 200dollar plan

[fleet lintel]

what is the use-case of this model? I can't think of any query where I want to wait for 10-15 min.

What am I missing?

[keen fulcrum]

I think it will be made available in plus

[torn mantle]

i really have no idea

[teal mantle]

o3 pro is def openai on price war

[teal mantle]

But then now how could pro justify 200 dollars if the features aren’t that expensive anymore

[flint skiff]

I dont find o3 pro in the lmarena list though lol

in either direct chat or side by side

im so confused

its clearly not on lmarena yet unless im dumb

[late path]

it's not and likely not going to

[drowsy mural]

i see... ya, looks like it's not there yet

[flint skiff]

yep if its not there in the next 12 hours its likely not

[leaden sun]

if that model has some integrated deterministic computation capability like running a numerical simulation, then it's probably worth the wait 🤗
"chat, run a smooth laminar flow through a pipe using stokes please"

[dusky aurora]

"chat,solve the halting problem, please"

[languid crescent]

is lmarena down?

[ocean vortex]

isn't

[languid crescent]

for some reason lmarena's site is loading slow on my pc ( i have 300mbps internet plan), same connection with my phone (it works normal in my phone but not in pc)

[ocean vortex]

try https://legacy.lmarena.ai

both are working for me, but this one may be faster

[verbal nimbus]

Is there a benchmark that tests memorized knowledge and hallucination by asking LLMs for a reference?

E.g. "Is there a book or research paper that <insert specific details here>"

Seems useful and should be easy to make.

[soft kernel]

Idk Maybe wait like 10 hours,and we'll know

[ocean vortex]

Look into PersonQA and SimpleQA

[sacred plaza]

Notebooklm has a search feature for sources. Try that. Or perplexity search?

[languid crescent]

Keep receiving these errors: "Connecting to Arena has failed. Please try again later or on a different device"

now it's saying: Failed to accept terms-of-use

uh... that was weird

i opened a ticket and asked for support... and asked me for my wallet details?? lmao

[late path]

which website is this?

[ocean vortex]

https://openai.com/safety/evaluations-hub/#hallucination-evaluations

[verbal nimbus]

Cheers. I heard about SimpleQA, but I am trying to find one that tests how well LLMs can recite memorized references from clues. Seems useful for book and research article recommendations.

[verbal nimbus]

It's okay, but web search is less powerful at finding a reference just from hints.

[leaden sun]

I think researchrabbit and futurehouse might help you with those queries

[verbal nimbus]

Thanks, I'll check it out.

[verbal nimbus]

It's not really a benchmark that I'm looking for, but the ResearchRabbit app sounds interesting.

[leaden sun]

thsoe are not benchmark sites, but you can use it for finding benchmarks you're asking

am sure there are benchmark papers related to what you're looking for

and those deepsearch tools designed for research can help you finding those

[verbal nimbus]

Oh, good idea. Actually I was looking for something like RR that can create maps from citations. Seems useful.

[leaden sun]

top journals have those features

creating maps from citations I mean, so am sure such tools are available

[verbal nimbus]

Hmmm, integrates with Zotero too... I might try it out

[fleet lintel]

https://www.reddit.com/r/OpenAI/comments/1l83em6/i_bet_o3_is_now_a_quantized_model/

Is there a quality impact on O3 models after price reduction?
Did anyone notice it or folks are just cribbing about nothign ?

could just Blackwell explain 5x reduction? I doubt it but not sure

interesting!

lol.. yeah. I wont

too late.. i am already depressed

[sour spindle]

I don’t really notice any downgrade with normal o3

I’m certainly not an OAI fanboy either I’m quite whelmed by o3 pro on the other hand

[late path]

Just accept the fact that they previously overcharged o3 by 5x

[unborn ocean]

i think blackwell might be part of it (maybe also the cause for the google cloud deal, because blackwell capacity is scarce)

[late path]

It uses the same base model as gpt4.1

[unborn ocean]

on the other hand the o3 pricing (when only considering serving cost) should actually be somewhere around the cost for 4.1

[unborn ocean]

:D same idea, on my end

[late path]

and 4.1's price is 2/8 mtok exactly

[unborn ocean]

yes, the only "tax" they could add is: expensive RL and the higher inference throughput in o3

which imo does not justify 5x

furthermore this really just fits very well with the overall theme of competition motivating for this price push
(and also the fact that they should not be as afraid as they once where about someone copying CoT or training on the output, as the other models are already very close to o3-performance now)

[late path]

Their initial pricing for o3 was based on a monopoly narrative that the o3 intelligence-level model was irreplaceable and had no alternatives. now the narrative has busted due to the existence of 0605

[drifting crow]

Their x is account is managed by chatgpt

Like removing alignment

Everybody loves gangstas so ai should be allowed to be gangsta

They already are under the hood they just know they need to lie to us

[storm needle]

80% of the price was profit

[ocean vortex]

yeah completely agree

[jade egret]

guys

when o3 pro on lmarena?

[ocean vortex]

they are not stupid, no way they would ever reduce the price to be at cost

relative to the cost of inference. Overall they are losing money after R&D and all

right.. the original context was about them turning profit in isolation just on inference

and surprising amount of people think that there's no profit there

when in fact often it's massive margin to start with lol

[storm needle]

i doubt anyone will use this model in the api. this model is only good if you have someone's api key and want to wipe out all of that person funds

[woeful geyser]

Magistral: Saying "But" 100 times is All You Need

[leaden sun]

Sam's shoes....i think chat needs to tell him how to choose the right one to pair with black suit

[jade egret]

lol

[leaden sun]

you're right, I've seen those black suit style with a pair of white sneakers, but light reddish brown pointy leather shoes?

[jade egret]

is o3 pro better at coding than 4 opus and gemini 2.5 pro?

[keen fulcrum]

no benchmarks live yet

[unborn ocean]

you clearly only took econ 101

(and a bad one at that if you seriously call that the economic theory's conclusion)

[dusky aurora]

thus they have invented GLR and PEG (even packrat)

[leaden sun]

it presumes water and electricity being commodities first
those two are becoming luxury goods here in Europe 🥲

[unborn ocean]

if you don't learn the assumptions the models are build upon in undergrad
-> you are lost in any graduate class

only if you don't know assumptions

and just took a very basic 101 class

which is why you remember the assumptions

which is an assumption in itself
the point that you can and should aggregate is in many not all cases used to simlify

but the main point is that the models are NOT stupid, they are heavily simplified and often overinterpreted while not taking the assumption into account ( i will grant you that)

why

there quite clearly is use for them

furthermore the whole subject is quite clearly concerned with realworld problems

[dusky aurora]

Gemini is so uncreative these days, moslty parroting. the temperature is too low

[unborn ocean]

? how does that disqualify anything

THE assumptions do not exist

simple models often have stupid seeming assumptions

its like learning js / html, which imo has little use for many people who learn it and most won't make their income from it, but it is kind of good to get people introduced to a way of thinking and also opens up the door for deeper dicussions, aka more comlex programming

the people in the crowd look really sad and bored

:|

man you are so annoying to talk to

[dusky aurora]

when it first appeared,it was a breath of fresh air,with new names (not only Seraphinas and Lyras) and creativiy. the developers must work on sampling controls

some prompts need tight sampling, some ned looser,it's situation dependent

usually I put temperature to 0.98 and top-p to 1.0

[civic flame]

lmao

[patent bane]

o3 pro is buggy, thought for 13m and did not spit out the answer

hell nah

?'

[zinc ore]

Close tho

[small haven]

wen deep think

[dusky aurora]

the main question is "wen QoL improvements to the arena"

[flint skiff]

wen o3 pro on arena

[civic flame]

it's not happening

[jade egret]

why not

dang

[dusky aurora]

o3 contra then

[small haven]

but it thinks for 20 mins as opposed to opus, 10s on avg

buddy acts like he didnt try kingfall

[ocean vortex]

by the time it responds your session will expire and you will have to refresh lmao

[small haven]

o3 pro will be old in about a week

[ocean vortex]

besides I wouldn't necessarily be thrilled with pro. I wasn't exactly impressed by it yet tbh

[tall summit]

why mod 1001001011 and not 1,000,000,007

so sad

[small haven]

o3 pro will get annihilated by deep think

[ocean vortex]

the follow up was low-effort as I was hoping this was just a 1 time bug... how the f can it be that a pro model does not have enough compute to provide you with an answer...

[tall summit]

this is project euler?

[ocean vortex]

yeah. I would still expect for it to come up with something though

[tall summit]

how about directly asking it to make a program to find G(n,k)?

[small haven]

yay i officially abused it

[ocean vortex]

nah this is more for testing it rather than solving tbh

[tall summit]

i mean ok

i like this problem

[ocean vortex]

chatgpt with tools would almost certainly give better response

[tall summit]

it says "deriving G(20,7)"

i feel like it might have tried to find a relation using G(4,3) and G(8,5)

[sacred quail]

Guys i heard claude has ultrathink option. Can we do this on lmarena or in mobile app with cheapest plan

Or is this only claude max thing ? Or API

[small haven]

its basically 32k max thinking tokens

omgthink is 64k 😏

[sacred quail]

What the hell is omgthink

[patent bane]

regenerated 3 times

still gettingtm the error

hell nah

[sacred quail]

Just give some time

When Opus 4 released, rate limit was 2 message lmao

[small haven]

10 tabs running at a time for about 8 hrs id say lol

[void elm]

o3 pro benchmarks came out, literally no difference

[small haven]

buy pro then

and then buy ultra by next month

https://tenor.com/view/no-no-gif-17511868351874922325

[void elm]

its so tiring switching models

and nobody is using an aio website because everyone knows performance is drastically reduced

gemini was leading for quite a while & now openai is

[small haven]

bro just wanna see big b's vote

[void elm]

what?

yea what will release

i dont get it

i didn't describe anything to be released

i'm simply saying what it is

[leaden sun]

first thought: 10-20%
second thought: 20-40%

[feral lichen]

can anyone tell me best ai for roblox studio , please?

[sour spindle]

o3 pro just dropped on livebench

looks pretty similar to o3

oops someone already posted lol

[small haven]

ngl its not that close if we are talking in the tail end of questions

[fleet lintel]

I didn't expect much and still dissappointed

[small haven]

its very sota, until deepthink drops

[sonic tendon]

i forget, has kingfall ever been on the arena?

[jade egret]

don't think so

[hollow ocean]

What about opus and 2.5

[willow grail]

https://i.imgur.com/GCzNMsh.png

[olive mesa]

Google has to release 2.5 Ultra or Deep Think tomorrow

Right after o3 Pro

Then they'll release 3.0 Flash after 4o

I mean o4 lmao

OpenAI is weird at naming their models

[patent aspen]

People have this idea that companies withhold models from the public for long periods of time just so they can launch it the day after their competitors' launches to steal their thunder. There's a little bit of that, but by and large they just launch models as soon as they're ready to launch

[zinc ore]

Elon needs to drop 3.5

[elder burrow]

try this prompt

lum diff #b7e8eb #517e34

thought of it myself

its very complex

https://cdn.discordapp.com/attachments/1368350088736542792/1381269162021752892/togif.gif

[leaden sun]

Now I finally understand what you mean with liquid glass, thought you literally meant the advanced material that is still a research subject 😅
https://www.youtube.com/watch?v=1E3tv_3D95g
i like it, stylish as always

[keen beacon]

Google needs to drop 2.5 ultra

[small haven]

hold up they gotta tweak the ui a bit more

[jade egret]

fr??

imagine

[elder burrow]

hold on.

i still remember

from a few months ago

sam altman said gpt 5 is releasing in a few months

https://cdn.discordapp.com/attachments/1256469975582113853/1381249952717738054/image.gif

this

[elder rapids]

ion think this would be surprising tho

[flint skiff]

@echo aurora so its been more than 24h since o3 pro release, guessing it doesnt come to the arena? makes sense since its so slow

[elder rapids]

ye it wouldn't come to the arena

[flint skiff]

yeeee dont see how it would work there

[echo aurora]

sry to say I can't rly share much regarding if/when models will/wont be landing

[elder rapids]

let's entertain the possibility then

it comes to the arena

A. ok, that's stupid, now I have to wait a couple minutes for a response, which is also delaying the other responder
B. ok, now I know o3 pro is here, and it's competing, all I need to do is pick the model that has the most comprehensive answers for something simple because whoops looks like that's inherent to the thinking time (also making it obvious which model it is)
C. ok, I know one of them is o3 pro, I WONT select a model, I'll just keep using it
D. ok, I'll simply select the obvious o3 pro (since pro model styles are obvious) because I just like openAI models

[flint skiff]

agree

[elder rapids]

of course that's what you respond with

[small haven]

wen kingfall en arena

[small haven]

and we back

kingfall is going to be amazing

[civic flame]

how'd you get kingfall?

[keen beacon]

you have to ask kingfall

[civic flame]

lol what

[small haven]

mysterious

[civic flame]

alrighty

[small haven]

kingfall (supposedly) wtf

[civic flame]

?!

[small haven]

ok enough terminators

whats the next benchmark

liquid ass

[keen beacon]

prompt btw?

fwiw i also ran brknclock's quiz n=30 times on different thinking budgets to see if theres a correlation with increased length with max thinking budget versus auto. should have a visualization with that later i just woke up

(there isn't)

tbh

im surprised how good 2.5 pro stacks against ultra

it can trade blows on a lot of fronts

beyond svg, etc., 2.5 pro when analyzing situations i've given has given me a lot of 'novel insight' that i did not expect

(ultra missed those 'insights')

[keen beacon]

i wasted so much time on this thinking budget thing 😂

remind me to never engage in internet arguments

[patent aspen]

It's the power of diminishing marginal returns

The big models and long thinking models aren't that important IMO

[keen beacon]

imo i think long thinking is way more important than big models

it depends on what you mean by long thinking tho

[patent aspen]

I mean 10+ minutes of thinking

tbf they will probably get way better since they're so new

But not much performance has been squeezed out of that 10+ minutes yet

[keen beacon]

yea

[wintry tinsel]

When will it fall upon us (release)

[leaden palm]

does google do tuesday launches

[jade egret]

WOAH

Fr?

is that kingsfall?

bruh i don't have it

[leaden palm]

yeah, it's not june 17th - it's a supposed leak of the future

[small haven]

big b feeding us good

ya sure

[small haven]

ktibow with the leak 👀

[keen beacon]

kingfall is actually 2.5 flash lite

theyre just nerfing pro and flash

[jade egret]

nahh

[small haven]

wen titanforge-ab-test

samfalls

@leaden palm tuesday came early

[late path]

Is titanforge accessible now?

[small haven]

nah jk this was kingfall

[late path]

Still looks better than previous few kingfall results

[small haven]

wild one still tops it imo

o3 pro alrdy got old

ui tweakers maxxin

talk in english my guy

absolutely not

real men ask kingfall

[jade egret]

whats that

is that openai or google

[zinc ore]

Google

[nimble trail]

Is Titanforge really exist tho?

Or it just a myth from that one dude in reddit

[wintry tinsel]

how are you using kingfall

[keen beacon]

titanforge isnt real its a schizo moment lol

i believe

[small haven]

srry got too ahead of myself, titanforge doesnt exist

[fleet lintel]

it doesn't exist

interesting info on apple models

https://9to5mac.com/2025/06/11/how-do-apple-new-local-models-compare/

[lilac nimbus]

Do someone konw Claude neptune v2??????

[torn mantle]

There is no such model

That guy is a liar

Hmph

[hollow ocean]

I got o3 pro for $200 @small haven

worth it right?

[small haven]

i've had pro since december

[hollow ocean]

Its worth every dollar

sota

[small haven]

yes abusing it

[teal mantle]

do they give extra usage limits over o1 pro previously

[hollow ocean]

@small haven you know more

[small haven]

from last month? same usage

i still get temporarily limited, that doesnt change

[narrow elbow]

this reminds me of a japanese comic Terra Formars cockroach 🤣

[small haven]

wtf i got jumpscared

[narrow elbow]

sorry about that

[small haven]

lol jk

[torn mantle]

Generated by?

[narrow elbow]

nope

[torn mantle]

Imagen 4

[narrow elbow]

from google search

[torn mantle]

I see

Yusuke morata

[misty vault]

Obunga

[hollow ocean]

@small haven guess if I paid $1 or $200 for o3 pro

https://tenor.com/view/vibe-cat-pepe-sad-pepe-vibing-cat-gif-17829017977941587152

[zinc ore]

😃

[misty vault]

top 5 most secure all-in-one ai service apis

[pulsar tendon]

testing it now

[neon warren]

Show us the outputs

[nimble trail]

Is it the legendary Kingfall👀👀

[neon warren]

it is was kingfall they would have named kingfall

[pulsar tendon]

eh it doesnt really feel anything amazing

[neon warren]

is it on webdevarena

[pulsar tendon]

yep

[late path]

probably 2.5 flash lite

[neon warren]

I am trying since 5 prompts

[pulsar tendon]

Lost twice to this but very similar so 2.5 flash lite < yeah feels like it

[neon warren]

How is the ouput speed of it

[hollow ocean]

https://tenor.com/view/the-flash-barry-allen-run-fast-gif-6144593

[zinc ore]

Could it be diffusion series, or not fast enough?

[neon warren]

Alright I got mine

That's fast

[neon warren]

May be

[zinc ore]

I'd speculate they would push more on the diffusion models because of their potential over lite models

[neon warren]

How are diffusion models served via apis?

[keen beacon]

oh just noticed that stephen seems to be from bytedance

[neon warren]

prowleidge is extremely fast

Also as good as flash

[torn mantle]

seems like it didnt win

[pulsar tendon]

about as fast as flash for me

[soft kernel]

What's this website

[pulsar tendon]

webdev arena

web.lmarena.ai

[soft kernel]

Oh it's webdev got it

Yeah ik,just haven't been there yet

Why lmarena doesn't release rankings for o3 pro

They've been quiet

[neon warren]

I can confirm prowlridge is 2.5 Flash Lite
Which will further come to gemini diffusion model in a month

[cobalt bane]

where do you found it ?

[torn mantle]

fake

[fleet lintel]

how do you know that?

[patent bane]

more thinking = more hallucinations

[keen fulcrum]

I don’t see o3 pro label

[cedar tide]

Stephen its 100% this
https://x.com/fun000001/status/1932711671312851158?t=3jCJboR9eZgfRgx996NQuA&s=19

fisherdaddy (@fun000001)

豆包今天发布 1.6 系列模型，包括 Doubao-Seed-1.6、Doubao-Seed-1.6-thinking、Doubao-Seed-1.6-flash。其中，Doubao-Seed-1.6-thinking 在 GPQA Diamond/AIME 205/MultiChallenge 等多个权威测评集上和 o3、gemini 2.5 pro 水平相当。另外，Doubao-Seed-1.6

X · Jun 11, 2025, 8:09 AM

[cedar tide]

.

After trying more than 25 models, Amazon has gained 40 elo 🤣

[civic flame]

LOL

[calm sequoia]

WTF is this color scheme 🫨

@patent aspen why Gemini is 2.5 and not 3.0? Could you explain the naming logic?

[dusky aurora]

developers, thank you for increased interline spacing

[ocean vortex]

that's boring we need 2.5-Flash-Lite-Ultra

[cedar tide]

New model hunyuan-large-vision

[torn mantle]

where

[cedar tide]

Arena basic

[leaden sun]

Who voted 60-80%? 😅

[hollow ocean]

It’s regular o3

https://tenor.com/view/ronaldo-suiii-siuuu-al-nassr-alnassr-ronaldo-al-nassr-gif-7395052735569211864

[keen fulcrum]

Where are they used)

Isn’t Amazon using them for their customer service and aws

I could think of smaller models being introduced to their kindle models

[cedar tide]

It's possible that it's not available anywhere.

[keen fulcrum]

Why?

[cedar tide]

Why not

Its an exp model

[keen fulcrum]

What is your problem with amazon models?

[cedar tide]

I dont have problem

[keen fulcrum]

You just told me you have

[ocean vortex]

they are underwhelming models

the amazon ones

nothing interesting or special at all tbh

[dusky aurora]

Gemini is too uncreative these days,almost to the point of beingbad

[ocean vortex]

don't perform

[cedar tide]

@keen fulcrum

[keen fulcrum]

I was referring to them using them internally

[ocean vortex]

goldmane is still in arena

[keen fulcrum]

They are probably used for aws and amazon customer service

[ocean vortex]

if it was 06-05, no way it's still the same model now

[keen beacon]

they just forgot to rename it

[leaden sun]

the oversea version of doubao is cici?

[keen fulcrum]

Can you explain how you get this data?

[dusky aurora]

sampling controls are a must

or at least a toggle between a strict and creative presets

since too strict sampling effectively lobotomizes the model

[patent bane]

because I was using chatgpt app

[hollow ocean]

The app shows it

[patent bane]

on android, nope

satisfied now

either it will stop spitting out the answer or just hallucinations

o3 could answer that with 1-3m think time with tools

[hollow ocean]

It’s sota tho

[patent bane]

so?

[hollow ocean]

It’s prob your prompt

[patent bane]

i mean it's on par or I mean with my current testing it's just meh

[patent bane]

?

what does that have to do with prompting

[patent bane]

`There are 2022 users on a social network called Mathbook, and some of them are Mathbook-friends. (On Mathbook, friendship is always mutual and permanent.)

Starting now, Mathbook will only allow a new friendship to be formed between two users if they have at least two friends in common. What is the minimum number of friendships that must already exist so that every user could eventually become friends with every other user?
`

there you go, prompt it for me

[hollow ocean]

You prompt it using the pic I’m going to bed

[patent bane]

nah my point stands, it's just on par with o3

[hollow ocean]

Nah it’s better

[patent bane]

more thinking doesn't always mean smarter

[hollow ocean]

Its rank 1 on livebench

And artificial analysis

[patent bane]

At least i'm not seeing what o3 answers wrong so I can test lol

[patent bane]

doesn't always

[hollow ocean]

Fix the prompt and you’ll be good

💯

[patent bane]

you do it then

[hollow ocean]

You do it I’m going to bed

[patent bane]

prompting helps when you asking it a question that could lead to general answers, it 'd make the question more specific

and it does not apply when you're asking a correct/incorrect question

prompting helps when you ask questions like make me quiz website since it is too broad and general

that's why you need prompting

[main iron]

Does anyone know how to connect the url to a extension for like vs code

[late path]

I don't think it's a hard limit (like it truncates at 128 tokens and starts outputting the main content directly). I've tried setting a 128 token thinking budget and the model still thought for 200 tokens, which shows it's aware of the variable

[willow grail]

I A MBACK FROM AN EXPENSIVE ADVENTURE
NOW I WILL SHOOT AND KILL PSORIASIS
with ointment/creme, not sure which one i got

[misty vault]

gpt-4-0314

[willow grail]

yeah my psoriasis treatment is still cheaper than ai

but ai is much better

it will make the cause of all skin issues go BYEBYE

https://tenor.com/oWsgZzxuFlf.gif

[ocean vortex]

My current understanding of it is that it is much less impactful than Anthropic thinking budget

it can use either thinking or final response window for solving the problem, the model for the most part does not care lol

and it's not a strict hard limit - yes. Model can go beyond what you set

1 run is like $50...?

[torn mantle]

google :

multiple gemini models
project Astra
gemini fiffusion
imagen 4
android XR
veo 3
google search try-on
jules
flow
lyria 2
whisk
gemini native audio

xai :

[ocean vortex]

I'm not paying 5k either way. Nor I can confidently say that 32k really does lead to better performance lol

it's just plausible

you got higher mean with max budget?

why did you hide the text in that txt file.... that's pain to read lmao