general

[keen fulcrum]

Which of these is better?

[drifting thorn]

o3 mini

2.0 Flash Thinking

[balmy mist]

wait so we get unlimited gens with veo in studio right?

[tall summit]

kinda rigged to bring in claude 3.7 given the list is mostly mini models

[oblique flint]

idk if o3 will beat it tbh. o3 might be smarter but most people might not like waiting long for an naswer

[balmy mist]

https://x.com/legit_api/status/1912498684337479686

ʟᴇɢɪᴛ (@legit_api) on X

o4-mini and o3 were removed 4 hours ago from here

this can mean anything but usually when they add models here, it's a day before release / same day.

removing such models COULD mean it's not coming today - but near enough so don't over react if so

X · Apr 16, 2025, 1:29 PM

[keen beacon]

yeah my source says today looks increasingly unlikely

[balmy mist]

that sucks

i hope r2 comes out today lol

or sum other model

[calm sequoia]

Likely to release or not release? And is this about o3 or o4-mini?

[keen beacon]

likely will not release today

bar a big surprise

it's about both

https://x.com/OpenAI/status/1912506271187832904

OpenAI (@OpenAI) on X

Livestream in o3 hours.

X · Apr 16, 2025, 2:00 PM

well

turns out they managed it

[balmy mist]

omggggg

omggg

ahhhh

i just jumped out my work meeting

when i saw the tweeet

i love open ai

i was going to be so sad today

[misty vault]

Please refrain from busting all over the LMArena community discord chat

[balmy mist]

ill try bro

[keen beacon]

might not be able to if it actually beats 2.5 pro in everything

[balmy mist]

lmaoo o3 hours

i love that

so no mini today?

[keen beacon]

they always sneak a hint for what they're announcing in the tweet announcing the stream

[keen beacon]

either today (as they have never not launched them at the same time) or it they're feeling quirky tomorrow

i think they just said o3 because it's easier to make it a pun or whatever

how else could they sneak in what they're announcing

[balmy mist]

o4 minus 1 hours

lmaoo jk

[keen beacon]

o4 min[i]-us 1 hours

[balmy mist]

you think they have a team dedicated just for puns?

i wonder how much they get paid

[keen beacon]

i am also told the benchmarks for o3 given in the last preview in december are now out of date

the model has improved since then

[keen beacon]

lmao

[balmy mist]

wow sama said they managed to make o3 way better which sounds like a reach, "improved since then" seems more reasonable

anyone have those benchmarks handy?

i wanna set expectations

wait @keen beacon you are talkign about jan benchmarks?

i thought they just teased o3 in december

and gave the actual benchmarks in jan or feb

[keen beacon]

huh

[balmy mist]

cant remember lol

[keen beacon]

no it was in december i swear

because it was part of a stream in the oai christmas run

[balmy mist]

yeah you are right

they just said they would release it in feb

[balmy mist]

if they improved on this then wow

and it got a high score on arch right? like it passed

[balmy mist]

so you are saying its better than this?

i cant believe that lmaoo, they might as well call it o3.1

[keen beacon]

i'm not sure about arc agi performance but i know it has improved performance on the other benchmarks

mainly down to going from 4o as base to 4.1

[balmy mist]

that makes sense

https://x.com/sama/status/1908167790336651720

Sam Altman (@sama) on X

we were able to really improve on what we previewed for o3 in many ways; i think people will be happy...

X · Apr 4, 2025, 2:40 PM

do you think december o3 is better than 2.5 pro?

what did gemini 2.5 pro get on arc and that math benchmark?

frontier math

[keen beacon]

perhaps just about

and not in everything

[balmy mist]

those are the only two benchmarks i care about and simplebench:
frontier math, arch1&2, Simple bench
we need a graph that combines them

[keen beacon]

they haven't run frontiermath on 2.5 pro iirc

[golden ocean]

this team consists of ai agents

[keen beacon]

something about rate limits

but 2.5 pro is very very good at maths

i'd say better than o1 or o3 mini or any other model currently available for that matter

[balmy mist]

that is odd

i cant wait to pay $200 again for SOTA

[thorny drum]

2.5 pro is so much better than any other model rn at math

wonder how it stacks against o3

[balmy mist]

not o3 technially

well december o3

[keen beacon]

i would expect 2+ point gains on most benchmarks vs december o3

[balmy mist]

that good enough for me

[thorny drum]

i mean idrc about the benchmarks its just about the cost

[keen beacon]

codeforces will be very interesting

[thorny drum]

december o3 low is like $200 per task on arc agi

[keen beacon]

i think it'll reach 3000+ with the updated base

because 4.1 is miles better than 4o at code tasks

same for swe bench

[balmy mist]

cause decem o3 prob was better than 2.5 pro, thats why oa is releasing it now instead of skipping them like they said they would

[thorny drum]

the model they release very well could be weaker than december o3

[balmy mist]

that competition coding right?

[balmy mist]

impossible in terms of benchmarks

we have the records lol

unless it cheaper to run, like way cheaper

then that is a win

[thorny drum]

i guess o3 low is only 2x the cost of o1 pro

[drifting thorn]

Just tried ChatGLM Z1 Rumination

[keen beacon]

yup

[drifting thorn]

It’s thought was good, but the base model itself is trash

[keen beacon]

it won't be

i can assure you

[thorny drum]

🤷‍♂️

[drifting thorn]

But I think those big techs should have a reference

[balmy mist]

why would they release it if its worse than december lmaoo

they just annouced they got the 40 bill funding

[drifting thorn]

Btw the “rumination” means extended “thinking” time and the ability to call tools multiple times while in its chain of thought

[keen beacon]

if it doesn't beat 2.5 pro in most benchmarks, they won't release it

[balmy mist]

fr

[keen beacon]

i heard (again from a source) that they delayed the model after 2.5 pro by a couple of weeks

because they wanted to make SURE it didn't make them look like they're behind

[balmy mist]

i hope this makes google release nw

i love this man

war of ai

[keen beacon]

i think the higher likelihood of a reaction

is from deepseek

with R2

which would be very cool

[balmy mist]

this needs to be a netflix doc

oh yeah

i would love that

imagine they release it during the livestream

lmaoo

[keen beacon]

i still think the most likely course of events is that R2 releases next week

[balmy mist]

but nahh

they gonna wait

[keen beacon]

2.5 flash, updated 2.5 pro tomorrow or friday if i had to bet

[balmy mist]

so they can steal iq

[keen beacon]

google and openai have a long running tradition of trying to steal each other's thunder

it used to be openai and anthropic but now anthropic is too far behind

[balmy mist]

yeah they got that aws partnership so they might gucci with that

it seems like its apple and oa vs google v anthopic and amazon, is that correct? not sure where meta fits in but meta vs deepseek(while also being against the other players?) lol

[ember rapids]

Nightwhisper tmrw?

[keen beacon]

doubt

[balmy mist]

not sure if microsoft still is buddy buddy with oa

[ember rapids]

Interesting how OAI chose to release today instead of Thursday

[keen beacon]

they're trying to make themselves more independent these days

they're failing quite hard mind you

[balmy mist]

nw tmw would be goated by google, but i dont think nw is better than o3 or o4 mini imo

[keen beacon]

but

[balmy mist]

@keen beacon what if nw was better? you think they release it?

[balmy mist]

cause i was gonna be sad, @keen beacon pulled in a favor for me 🙂

[keen beacon]

google are less willing to sit on things than oai are

if they do have a better model and it's almost ready, they'll push to have it out by end of next week max

if they don't expect to be waiting a couple weeks or more

[balmy mist]

bet, these are exciting times man, im just happy google stepped up their game

it was looking scary a few years ago lol

[keen beacon]

it's so much better when openai have pressure piled on them

to the moon 🙏

[alpine coral]

i would expect o3 to outperform gem pro 2.5 on most benchmarks; question will be though, by how much and what cost?

like if it's marginally better but twice as expensive, gemini would still be ahead imo

but if it blow gem 2.5 out of the water, then yeah who cares (for now anyway) what it costs ha

[fleet lintel]

did they even launch o3 on lmarena?? probably not

[balmy mist]

yeah gemini 2.5 pro advantage is that it is free and cheap and SOTA

[thorny drum]

its gonna be way more than 2x as expensive

[balmy mist]

they are most likely still going to be SOTA bc of that, but in terms of intelligence i bet on OA, i think thats their goal with these models today

[alpine coral]

well, kinda.. if the discord server counts ha
(or perhaps it's o4-mini) private model

[thorny drum]

like waaay more

[balmy mist]

they already released their cheap models on monday

today is not that day

[alpine coral]

yeah i would expect it'll be more

[balmy mist]

idc about price

im here for the most capable model and thats why we like OA

we wanna see reasoning breakthroughs

[keen beacon]

nope

[thorny drum]

o3 low was roughly 2x as expensive as o1 pro which is 60x more expensive than 2.5 pro

[keen beacon]

they don't ever test o-series models on the arena

i have access because i help with security and jailbreak testing

[fleet lintel]

2x expensive but significantly better is fine. but 5x expensive and marginally better is not good

[keen beacon]

that's through an oai-controlled platform so

[oblique flint]

1k / 1M output tokens incoming

[alpine coral]

i suspected it was red teaming ha

[balmy mist]

thats why you buy the subscription

[fleet lintel]

give something juicy to us

[balmy mist]

using api for o3 is nuts

or o1

[keen beacon]

lmao

like? 😭

[thorny drum]

so a baseline is 120x more expensive than 2.5 pro (on low reasoning effort)

i dont think 2x is even in the ballpark

or 5x

[keen beacon]

oh yeah i think you guys will be annoyed about pricing

i don't have official numbers but i have talked with a few people

[balmy mist]

idc about pricing, ik what to expect

[fleet lintel]

what to expect??? tell us.

[balmy mist]

just give me model!!

that you cant afford the api costs lol

and you need to sub up

thats what imma do

[keen beacon]

better than 2.5 pro in basically every regard but web development, however quite a lot more expensive

will be SOTA on basically all benchmarks

[balmy mist]

damn

so nw is still king

😦

[fleet lintel]

cool. and how much better on benchmarks?

[sage raptor]

in other coding tasks is SOTA ?

[fleet lintel]

marginal or significant?

[balmy mist]

how is it not as good in web dev tasks? thats weird, i think that has to do with system prompts and tool calling

you make any model better at web dev with clever prompting

[keen beacon]

it won't be tiny amounts better, fairly significant. but still in striking distance for deepmind

[fleet lintel]

2.5 pro is so good .. i am excited to see what o3 has to offer

[keen beacon]

swe bench, yes

arc, yes

codeforces, yes

i also hear it "did well" on aider polygot but no figures there

[sage raptor]

looks promising

[balmy mist]

who is noam?
https://x.com/legit_api/status/1912516320966430928

ʟᴇɢɪᴛ (@legit_api) on X

Noam's excited therefore I am

https://t.co/GFB7rf1gVR

X · Apr 16, 2025, 2:39 PM

@keen beacon he valid?

yupp

[keen fulcrum]

Here is a surprise for you:

[balmy mist]

today is a holiday

ewww

grok

[fleet lintel]

hard pass

[keen fulcrum]

They should offer API subscriptions, unfortunately they don't

[balmy mist]

hmmm, that kind of defeats the purpose a lil tho

[keen fulcrum]

Where?

[balmy mist]

but i see what you mean

[brittle tiger]

@keen beacon do you think we get o3 today if 2.5 capabilities didn't surprise ppl?

[keen fulcrum]

Thats PAYG

[keen beacon]

he's a big contributor in reasoning model research @ OAI

[balmy mist]

wow, im excited af

[keen beacon]

wdym?

[keen beacon]

goals?

[balmy mist]

lmaoo perplexity has not been the same since 2.5 pro came out tbh

[keen fulcrum]

So an AI engineer hopping around

[keen beacon]

taste testing every lab

lil bit of this, lil bit of that

[sick mountain]

what does member of technical staff mean

[balmy mist]

lmaoo smart man

[brittle tiger]

after gemini dropped sama tweeted "change of plans: we are going to release o3 and o4-mini after all, probably in a couple of weeks, and then do GPT-5 in a few months." was just curious if that was in reaction. my gut says likely but curious on your opinion

[keen beacon]

ah

yeah that wasn't the plan to start with

2.5 pro served as a reminder they weren't invincible

[balmy mist]

nothing was the same since 2.5 pro

pre 2.5 pro and post

its like o1 was SSJ and o3 is beyond SSJ lol

[keen beacon]

i also hear that although there is quite a bit of overlap there is some "healthy but fierce" competition between the team focused on o3 and the one focused on o4 mini lmao

[novel flame]

Check the Mixture of a Million Experts paper from last summer, it's pretty wild: https://arxiv.org/abs/2407.04153

[keen beacon]

lmao

[balmy mist]

i love it

[keen beacon]

https://x.com/kalinowski007/status/1912506846524633238 member of technical staff @ oai

Caitlin Kalinowski 🇺🇸 (@kalinowski007) on X

This is a big deal.

X · Apr 16, 2025, 2:02 PM

lots of hype posting going on

[balmy mist]

omggg its agi

we made it guys

lock in!

[keen beacon]

i think this is the first model within "striking distance" of AGI if you will

[balmy mist]

the hype is real man

who gonna livestream it here?

[keen beacon]

i think with the current rate of progress AGI is looking like it'll arrive by the end of next year max

[balmy mist]

im scared

what should I do to cope?

[fleet lintel]

i am sure that it is going to be SOTA but i am concerned about cost.

[drifting thorn]

Just when API for LLMs are available, I remember it costs several dollars per thousand tokens

[keen beacon]

ask chatgpt 😉

[novel flame]

AGI won't come from (autoregressive Transformer based) LLMs though. LLMs might help accelerate the research and coding though.

[fleet lintel]

hopefully not till 2030. companies wont think twice before firing us

[keen beacon]

lol

[keen beacon]

got some wisdom straight from our boy o3

[balmy mist]

wait you still have access lmaoo

[sage raptor]

livestream in o2 hours

[keen beacon]

yeah they don't deprecate the model previews until like a week after they're publicly launched

[oblique flint]

well, I would suggest checking out claudeplayspokemon and gemini plays pokemon lol. It's not AGI if it can't play a children's game

[keen beacon]

don't need to worry about payment just yet

🙄

[balmy mist]

limit doomscrolling lol

[keen beacon]

a matter of months ago the best llm couldn't get past even the first part

[oblique flint]

an llm beating elden ring is my AGI definition

[keen beacon]

just need to put things into perspective

[balmy mist]

damn so are you even excited?

[keen beacon]

well yeah

because i don't have the "-high" variants nor do i think i have o4 mini

[drifting thorn]

Is that u?

[balmy mist]

wow

[keen beacon]

lol no it's a linkedin profile i found scrolling through suggested connections

it would be my dream though 👀

labs are getting relentless trying to secure the best talent

[drifting thorn]

Ok it verified my guess on “two teams competing in OpenAI”

[keen beacon]

so much so that deepmind are paying people to sit around and do nothing because it prevents them from being poached when they might be useful later on

[balmy mist]

its crazy how much hype they are giving it, like i kinda dont know what to do

i wonder what google is thining

thinking*

[drifting thorn]

I’ve seen this this afternoon and I can hardly get what it means

[keen beacon]

i think they're probably still quite confident they can retake the lead in the next month or two

deepmind are the lab that i think moves the fastest

[barren prairie]

Week **

[keen beacon]

mostly down to the fact they have so much money to work with and their compute is unmatched

[balmy mist]

we need a chart for times when certain companies or models are leading, so we can see how long each company has held that title

[narrow elbow]

need another 12 days of live streaming non-stop,then google popping in, just like last time hahhaha.

[balmy mist]

i agree

OA the king of hype

[keen beacon]

iirc this arena has a video of how the leaderboard has progressed since it was created

which kinda works as that

[balmy mist]

oh wow, nice

[keen beacon]

best way to know if google are about to drop is if logan posts "Gemini" on twitter

he always posts just that word the day before a launch

if they're releasing a new gemma model he'll post "Gemma"

ain't no hypeposting around here kids

[fleet lintel]

google sucks at marketing

[balmy mist]

they lowkey dont have to market

they got so much money

[plain zinc]

It's much better than if they relied more on marketing.

Remember how they tricked everyone with Gemini 1.0 Ultra

[keen beacon]

if google invested as much into marketing as they did into research they'd be right on openai's ass

[plain zinc]

It's all because of dumb marketing.

[keen beacon]

openai are basically the only lab that actually know how to market

perhaps unfairly given the huge advantage they got with chatgpt's viral moment

but all the same

anthropic marketing sucks

[fleet lintel]

dont remind me of 1.0 models. they were horrendous

[keen beacon]

google's marketing barely exists

[drifting thorn]

LLM playing Minecraft is my definition

[keen beacon]

deepmind didn't really market r1 but it was carried by the press and social media

[keen beacon]

don't slander my boy 1.0 ultra... he was good at creativity. yes that was about it but 🥹

[keen fulcrum]

As Google is taking AI finally mandatory, we will see Gemini models absolutely dominating the next years

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

[keen beacon]

yeah google's TPUs given them a big advantage

[balmy mist]

deepseek doesnt really market but they market through their tech, thats the kinda marketing you really want, its so good that it markets itself

[keen beacon]

yeah i think that's what GDM are trying to do

[fleet lintel]

you mean deepseek. dont be too sure about it. china has huge power over media houses and they have incentive to hype it up

[keen beacon]

but thus far it has been less successful

[balmy mist]

no deepmind

[balmy mist]

^

[keen beacon]

if i asked almost any of my irl friends if they knew what gemini 2.5 pro was they'd ask me wtf im talking about

[keen beacon]

yeah my bad

lmao

yeah i know they have incentive

but most of the heavy lifting, at least with R1, was natural because of the model's strengths and it challenging american dominance rather than down to deepseek's efforts themselves

[drifting thorn]

True, media house in China hype Chinese AI way more than others

[keen beacon]

well, as you'd expect

[drifting thorn]

They hyped Deepseek R1 and sell sessions that “make you earn from Deepseek”

Which is a scam

[balmy mist]

china seems really good at marketing

[keen beacon]

i wonder if R2 will cause the same absolute hype storm that R1 did

[balmy mist]

like generating hype or commotion

manus, deepseek etc..

[drifting thorn]

Most critics in China is overhyping R1, claiming that it is better than o3 mini

[keen beacon]

you literally couldn't use R1 via almost any API nor via deepseek's own platform half the time because every gpu assigned to it was vapourised

[balmy mist]

wild

[drifting thorn]

Ofc

[keen beacon]

hopefully they've scaled up enough since then to be prepared for large load with r2

[balmy mist]

r1 was trained on o1 outputs right? so that means they cant release r2 until o3 is released cause training on o3 mini is not good enough

[thorny drum]

r1 was marketed by people losing money in nvda

the deepseek team didnt do any marketing really

[drifting thorn]

They were banned from buying Nvidia GPUs

[balmy mist]

thats why this moat crap is silly

[drifting thorn]

That’s why their server overloaded

[keen beacon]

they stockpiled tf out of them before the ban went into force though as the chinese do with most things

[drifting thorn]

Nah it was trained by R1-zero’s output

[balmy mist]

r1 was trained on R1-zero output or you are saying r2 is?

[drifting thorn]

R1-Zero is their internal model before public release of R1

As stated in their paper

[balmy mist]

ahh okay, they what was all that commotiion about openai getting made at deepseek?

was it jsut because they did it cheaper?

[drifting thorn]

I wonder when they can train new V3 based on R1, I think they will train new R1 based on new V3

And it will become master of hallucinations

[drifting thorn]

They didn’t state how V3 was trained…

I mean old V3

[balmy mist]

love this channel:
https://www.youtube.com/watch?v=yPxavsb2rgk&ab_channel=AIExplained

[keen beacon]

ai explained is pretty good

[drifting thorn]

Waiting for any LMMs to call out video generating models to have deeper understanding to the context

Thx for recommendation

[blazing rune]

GPT-4.1 made this 1st try in Windsurf: https://pong-html-js-responsive.windsurf.build/
Prompt: ```
Create a web-based Pong game using HTML, CSS, and JavaScript with these features:

• Player controls using W (up) and S (down) keys
• Computer opponent with beatable AI
• Score tracking for both players
• Game over when player loses by 10 points
• Pause functionality with spacebar
• Restart option with R key
• Clean, responsive design

[keen beacon]

https://x.com/ananyaku/status/1912523195175039398 from oai researcher - confirmed we're also getting o4 mini today. cc @sonic tendon

Ananya Kumar (@ananyaku) on X

Had a good breakfast today!

X · Apr 16, 2025, 3:07 PM

[patent aspen]

The only problem is that big AI breakthroughs tend to be bottlenecked on top 20-40 researchers at any given org

[drifting thorn]

Actually, GLM has a great solution(extremely large amount of thinking tokens with the ability to call tools multiple times in the inference time for 1 response)

[sonic tendon]

thx

[drifting thorn]

So sad that its base model is 32B dumb

[balmy mist]

i love that

4 mini strawberries lmaoo

but why are the strawberries smaller and smaller

[sonic tendon]

probably a coincidence? we shall see

[balmy mist]

red you streaming this time?

we need you

[sonic tendon]

yea, should be

[keen beacon]

just the 4o image gen being silly probably

[sonic tendon]

ok yeah i am

oh, for some reason i thought he'd have used real strawberries

[drifting thorn]

No way it’s real strawberries

[sonic tendon]

i mean

the occasion probably warrants it

[north vale]

it would just look like a worse image

[keen beacon]

lol i asked o3 for an SNL cold open after trump's tariffs

pretty funny

[tawdry meteor]

oh so they just come onto the arena after being cleared on safety and fully released?

[keen beacon]

yup

[tawdry meteor]

makes sense. excited to see the benchmarks

[balmy mist]

put into kling and lets start cooking

[tawdry meteor]

it's 1pm the livestream right?

[balmy mist]

est

imagine sama not on stream

[tawdry meteor]

the only real timezone /s

[keen fulcrum]

https://packaged-media.redd.it/h9c15ev2y4ve1/pb/m2-res_480p.mp4?m=DASHPlaylist.mpd&v=1&e=1744826400&s=7972827f4cd44f02d0a3d8e7cde0aa6cf004b630
This is apparently a helicopter

[keen beacon]

agi cancelled

[balmy mist]

@keen beacon https://x.com/TheXeophon/status/1912523048411951223

Xeophon (@TheXeophon) on X

🚨 EXCLUSIVE: Evals for the upcoming o3 and o4 leaked, straight from the presentation!!!!

X · Apr 16, 2025, 3:06 PM

[keen beacon]

there is no agi without twinks

[balmy mist]

is this true?

[keen beacon]

can confirm

[balmy mist]

damn

thats wild

the shapes are interesting

[keen beacon]

they say o5 will be an irregular shape

[torn mantle]

hes trolling

xd

[keen beacon]

truly groundbreaking

yeah ☠️😭

[torn mantle]

idk if this o3 model will be good

well i wont be able to use it anyways

[balmy mist]

maybe this means o4 mini will be faster than o3 mini? since it has a smaller strawberry?

[balmy mist]

bro o3 will change out lives

[tawdry meteor]

btw pretty sure R2 won't release until May

I'll link the source I was reading about it

[balmy mist]

that was the original plan but ppl said they updated that

[tawdry meteor]

🤔 when did they update that? I guess what I was reading was two weeks old

[keen beacon]

it's satire i wouldn't read into it too much

[tawdry meteor]

they said they wouldn't release this month

[balmy mist]

or sama's plan?

[keen beacon]

we're all going schizo

[balmy mist]

lmaooo

[keen beacon]

iirc i thought the expectation was R2 by end of April?

[keen beacon]

you're a little late

[balmy mist]

imagine the speaker in the livestream is o3 or o4 mini

they still didnt give voice to o1 right?

guess thats hard with reasoning

[keen beacon]

it's not multimodal with output like 4o is

id expect to see those capabilities built into gpt-5

or some of them

[balmy mist]

i really wonder how the inference will be for gpt5

[tawdry meteor]

South China Morning Post article from April 6 referenced a message from a business manager at Deepseek to clients that R2 would not be coming out March/April

[torn mantle]

it did? xd

[tawdry meteor]

But maybe end of April/beginning May doesn't count in that

The meaning was a bit difficult to parse

[keen beacon]

i kinda feel like their deadlines are fluid asf

they have a lot of pressure being put on them now

[tawdry meteor]

Yeah fair

[keen beacon]

they abandoned the date for R2 previously in favour of just "ASAP" when o3 mini dropped

[sonic tendon]

yeah

DS before may would be interesting - not sure about the odds there

*may

[keen beacon]

before may

yeah

i mean if it isn't april it will be the first half of may max

[sonic tendon]

whar?

[keen beacon]

like if they don't launch in april

[sonic tendon]

ohhh

[keen beacon]

i highly doubt they will leave it any more than another 2 weeks

[sonic tendon]

asdfhjlkasdfhjkl i thought you were just poking fun

[balmy mist]

wait so o3 is a researcher now?

[keen beacon]

it will be more agentic if you want to put it that way

rumour has it there will be updates to deep research this week too

[balmy mist]

@keen beacon im starting to see that 20k number really show up more

are they really doing that?

[keen beacon]

it's in the roadmap but not concrete

[balmy mist]

okay time to take out a loan

[keen beacon]

aimed at enterprise of course

if they tried to aim that at consumers i think sama would have to make sure there are no luigis around

[balmy mist]

im starting to see where openai is going, they really are pushing this product stuff

[sonic tendon]

VC money running dry

[keen beacon]

quick a16z and sequoia

bankroll sama

[balmy mist]

which makes sense, there is no moat in models anymore

[keen fulcrum]

https://www.warp.dev/pricing
this is honestly the best AI sub you can get

[balmy mist]

with a good product built around your model you do a lot

[keen fulcrum]

You can use that even for coding

[balmy mist]

yoo they better not blue ball me again i swear

[keen beacon]

https://x.com/i/lists/1676646159539130369 just take a look at this feed of oai employee twitter accounts

@altryne/OpenAI folks / X

X

[keen beacon]

nah the models are everything

what a troll

[sonic tendon]

what does that even mean

[keen beacon]

patience chair its a meme

[keen beacon]

the patience chair silly

[balmy mist]

but the models can be copied so it a given that everyone will have a good model

especially when the percent changes are so small

normies dont even notice the changes now

[keen beacon]

need me some of these chairs

cute car too

[balmy mist]

most of my friends cant tell the difference between gpt4 and gemini2.5

[keen beacon]

other frontier labs can copy each other, but if you arent a frontier lab and youre building products off of them i think ur gonna fail hard eventually

this is interesting

claude ever the yapper

[fleet lintel]

Best quality is almost infinite Gemini 2.5 Pro usage for Free.

People dont know but Google launched Gemini code assist and giving 2.5 pro for coding for free . I dont know why Google is doing it ?

[balmy mist]

you have a point bc before sama said to build with the the fact that models will get better overtime, but now openai is a product company and so is other frontier labs, where they are building tools around their own models and integrating it with stuff, so im not sure anymore

[balmy mist]

hmm explain this to me please, im kinda slow

[keen beacon]

they have to keep themselves afloat especially right now with how competitive things are

[balmy mist]

so its reasoning more for each output?

which is bad right?

[keen beacon]

it depends

[balmy mist]

if another model can get the same answer with less thinking tokens?

[keen beacon]

yeah

the less tokens it can spend reasoning to get the right answer the better

the best models will be able to most intelligently decide how much to reason

[balmy mist]

why did they color code it like that its so hard to see, but the thinking tokens are on the bottom right?

[keen beacon]

hm?

[balmy mist]

pretty much what gpt5 should be able to do flawlessly?

[keen beacon]

well it should get relatively close

needs to get to a point where you don't need any input minus your prompt to get the best answer

[balmy mist]

for this chart who scores the best?

[keen beacon]

2.5 pro iirc

obviously its sota rn

[keen beacon]

here's the reference

[alpine coral]

most of those tokens on the 3.7-thinking are often wasted.. it's 'reasoning' yields the same and sometimes inferior to just what the 3.7 vanilla model produces

[keen beacon]

i think it's o3 mini efficiency wise

[balmy mist]

wow

2.5 is really good

oh you are right

its second

and less thinking tokens

wait no

less thinking tokens

but more output tokens

[keen beacon]

but o3 mini high is also based on 4o mini 🤣

2.5 pro is the most yappy when it comes to output tokens

[alpine coral]

they're the same functionally

output/thinking tokens

whether thinking is abstracted away (or hidden entirely)

it's still completion being
<reasoning>
<answer>

rather than just
answer

[balmy mist]

https://www.youtube.com/watch?v=sq8GBPUb3rk

sama not there wtf

[keen beacon]

there it is

[balmy mist]

Join Greg Brockman, Mark Chen, Eric Mitchell, Brandon McKinzie, Wenda Zhou, Fouad Matin, Michael Bolin and Ananya Kumar as they introduce and demo the new o-series models.

[keen beacon]

AGI CANCELLED

[drifting thorn]

It’s the only model that will give me 10000 words sequel from my unfinished novel

[balmy mist]

but greg is there

[drifting thorn]

Which, it’s good to be yappy, but the bad thing is that it will often hallucinate when it has a long output

[balmy mist]

why sama

why

[keen beacon]

jokes aside

[balmy mist]

why does this not keep you up at night!

[keen beacon]

there are a lot of people there

[drifting thorn]

And forgets all the details I said before

[calm sequoia]

Smells like brute force search

[keen beacon]

this is the most people i've seen attending a stream since the 4o launch

[balmy mist]

sama only cares about memory

yeah wild

[keen beacon]

for o1 and o3 mini they only had like

3 people

[balmy mist]

what if sama is replaced by o4

and they are annoucing that today

[keen beacon]

this time they have almost 3 times that

how are they gonna be sitting

is it gonna be like

a long table with greg as king

or what

[balmy mist]

what if a new room?

something is fishy

there is a disturbance in the force

both

[keen beacon]

why are reddit mods such damn haters

no comment saying why, no message

not a dupe

god

[balmy mist]

wait what does this mean?

[balmy mist]

who is shrek and donkey

[keen beacon]

he posted a reddit post and it got removed

^

yes

check the live stream 🤣

[drifting thorn]

The moderators removed his post

[keen beacon]

#general message

[drifting thorn]

Gotta sleep rn

[balmy mist]

lmaoo

[keen beacon]

and the livestream yeah

ikr

what's he up to

[balmy mist]

he is a very busy man

[keen beacon]

he hasn't been at a launch stream for like

3 successive launches now

😔

[balmy mist]

he might pop in at the end

[keen beacon]

doesnt he have a baby tho i guess his schedule is wonky

[keen beacon]

busy enough to not attend a major launch but free enough to tweet all the time

smh

[keen beacon]

oh yeah good point

[balmy mist]

sama's plan

he cooking something

[keen beacon]

idk what to expect with o4 mini tbh

[tawdry meteor]

well I'm curious to see if o3 is actually better than g2.5pro, it's just so far ahead of everything else still

[keen beacon]

this guy is part of the agents research team @ openai

so i think we're getting agent updates or a new agent related feature

(alongside the models)

[drifting thorn]

And Sam Altman will be appearing on the livestream of new agent related feature?

[keen beacon]

it's part of the same stream as the one for o3 and o4 mini

so no

my bet is just on deep research upgrades

since gemini deep research made them look bad

gotta reclaim the lead

[alpine coral]

i actually think their existing one is superior to gemini's even with 2.5

problem is you get 10/month

[keen beacon]

what the 🤣

[tawdry meteor]

gemini is defo worse at news research imo

[tall summit]

oai researcher fanfic

[tawdry meteor]

but I haven't tested much on deep research for studies

[balmy mist]

its crazy this all started because sama and elon wanted to stop google from getting agi lol

[tawdry meteor]

and now they're suing each other and are gonna let google get agi lmao

[balmy mist]

i wonder what ilya is up to with SSI

[novel flame]

Crazier that things went sour because Sam wanted profits and Elon wanted fascism.

[balmy mist]

wild bro, kinda sad tbh

i wonder if they could team back up in the future

grok and gpt

vs gemini

ai wars

what i want is an arena battlefield for these models using agents powered by their models in a fight

[narrow elbow]

where is microsoft?

[balmy mist]

who?

jk lmaoo

they mia bro, they updated copilot

[narrow elbow]

hahha

[balmy mist]

thats about it

[keen beacon]

i dont think ms is trying to compete at all in the frontier space as themselves

[balmy mist]

who want to make this arena battle field with me?

[novel flame]

Either going down novel but ultimately wrong paths to AGI, or more of the same (autoregressive Transformer LLMs with more compute). I would love it if Ilya actually created AGI but so far all we know for sure is he’s really good at building yappy chatbots.

[balmy mist]

it can be a minecraft thing where each model gotta build their own castle and base and then attack the other ones with strategies etc...

[keen beacon]

that would be cool ngl

[balmy mist]

you think ilya is on the wrong path?

[keen beacon]

the run-down:

Greg Brockman - President, Co-founder
Mark Chen - Chief Research Officer
Eric Mitchell - O-series Research, Deep Research Core Contributor
Brandon McKinzie - Research/Member of Technical Staff
Wenda Zhou - Research/Member of Technical Staff, o1 Contributor
Fouad Martin - Agent & Systems Research
Michael Bolin - Research/Member of Technical Staff
Ananya Kumar - Research Lead, Core Contributor for o1 and GPT-4.5

[balmy mist]

there is a rumor that he saw something

[novel flame]

What you’re describing is WW3

[balmy mist]

i would pay for that tbh

like that could be the new benchmarks

[alpine coral]

agreed (they've chosen instead to invest $13bn into openai for that)

[ocean plume]

anyone test code claude 3.7 thinking vs 2.5 gemini pro and o3 high ?

[balmy mist]

yeah 2.5 pro is daddy

[ocean plume]

what better code and less bug

[sonic tendon]

who's ready for DeepSeek to overtake them in 12 days

[keen beacon]

they shouldnt have released 4.5 tbh

[ocean plume]

so who is grandpa

[sonic tendon]

at least in lmarena, they did last time lol

[torn mantle]

me

memememe

[sonic tendon]

for o1, i believe it did

[torn mantle]

they did on o1

[keen beacon]

i mean openai had o3 already at the time

if openai didn't stall every release they'd be ahead of other labs most of the time

but they sit on their best stuff a lot

[balmy mist]

is o5 being trained? or is focus on gpt5?

[keen beacon]

parallel teams

[sonic tendon]

the clone of terence tao i have locked up in my basement

[keen beacon]

gpt-5 already firmly underway, o5 in early stages

[balmy mist]

wow

[keen beacon]

now that the o3 and o4 mini teams have wrapped up they're being moved to mostly o4 and o5

[balmy mist]

why dont you go work for them?

[sonic tendon]

isn't o4-mini just a direct distill of o4?

[keen beacon]

i'm happy with what i'm doing now

o4 is not ready enough to do that i think

[keen beacon]

i don't think so

the o3 we're getting today is basically a whole different model to the one they announced

[balmy mist]

really?

what is it then?

[alpine coral]

yeah they said in that interview that they started training gpt-4.5 two years ago (and planning for it a year before that)

[novel flame]

I think Ilya spent a lot of time at OpenAI chasing bigger Transformers convinced that the bitter lesson / scaling laws would mean that GPT-5 would become AGI by virtue of its size alone. And I think he missed a lot of opportunities (R1, test time compute) that other labs spotted, and that a lot of the big ideas he used came from DeepMind papers. I think Ilya is brilliant, but he has chased Transformers for so long I’m not convinced he can let them go.

[keen beacon]

that is what i have heard from inside

they kept delaying it, then the name was changed to 4.5 and they decided to just "bite the bullet" because they didn't want to waste all the resources they had put into it and not even let it see the light of day

ilya leaving made things quite a bit worse

damaged morale as well

[keen beacon]

i dont think so lol. i think he saw how test time compute would pan out

[balmy mist]

hmm, we will have to see, i still think he has found a trick that most people are missing

[keen beacon]

yea i think

[balmy mist]

yeah i thought so

[keen beacon]

yup

[sonic tendon]

honestly, maybe they should've just scrapped it lol

[sonic tendon]

hmm

[balmy mist]

i think we can give him grace to cook, there is no way he not gonna produce soemthing goood

[keen beacon]

to be fair if i had put upwards of $150M into one model i'd literally just drop it out of principle 🙏 😭

[balmy mist]

if deepseek can do what they did, i know ilya gonna come with heat

[sonic tendon]

gpt4.5 sort of makes me doubt that scaling laws will actually hold up, no?

[sonic tendon]

rip

[keen beacon]

well of course lmao

oh wait

no

misread

it did not cost close to a billion no

it was still one of the most expensive training runs ever though, perhaps the most

[balmy mist]

you was right:
https://x.com/btibor91/status/1912544591821132135

Tibor Blaho (@btibor91) on X

Fouad Martin - Agent & Systems Research

X · Apr 16, 2025, 4:32 PM

[thorny drum]

more expensive than that grok one?

[calm sequoia]

Leo, why do you have so much info? This can't be open source material?

[keen beacon]

yes

although it's close iirc

grok 4 will probably beat it

[keen beacon]

i suppose you could say i am relatively well connected

~900B

3T+

lmao

[torn mantle]

he has intels

he know some openai staff

[balmy mist]

how many parameters is o1? and o3?

[torn mantle]

some of his friends

[keen beacon]

same as 4o lol

[torn mantle]

too

[keen beacon]

o1 is quite small

200b

yeah

200b

[balmy mist]

wow

[keen beacon]

beat me to it

[balmy mist]

see

thats why the strawberries are smaller

sama's plan

[oblique flint]

if o1 is so small, why is it so hella expneisve?

[keen beacon]

margins

[torn mantle]

profit margin

they want to earn a lot

[keen beacon]

semianalysis had an article on this, its the kv seq length

but yes they also have margins

[torn mantle]

but then

r1

[oblique flint]

do you have the link by any chance? Would be interested to read it

[torn mantle]

messed things up for them

[calm sequoia]

Only oAI lab?

[balmy mist]

i predict o4 will be 50 parameters

[torn mantle]

50 what

[balmy mist]

sama told me in a dream

50 b

[sonic tendon]

50 bdozen

[alpine coral]

lmao

[balmy mist]

sama told me to trust the process

[keen beacon]

can confirm this is NOT the case gang 😭

[keen beacon]

the majority is from openai but there are some scattered among the other ones

[keen beacon]

its been a while might be this one: https://semianalysis.com/2024/12/11/scaling-laws-o1-pro-architecture-reasoning-training-infrastructure-orion-and-claude-3-5-opus-failures/

be wary that some of the article isnt true

claude 3.5 opus/bigger model wasnt used in the dev of sonnet 3.5 according to dario

sonnet 3.5 was truly a generational training run

i wonder what anthropic researchers were thinking when they properly put it through its paces and it was cooking that hard

i dont think they expected it to be that good

it just lined up

[balmy mist]

their ceo used to work at google right?

[keen beacon]

all of anthropic's founding team were ex-deepmind iirc

[balmy mist]

lol

[keen beacon]

sorry no

dario was ex oai

[balmy mist]

its like everyone went against google, then went against oa

[keen beacon]

it was a mish mash of ex deepmind and ex oai

[balmy mist]

ilya was google at first right?

[keen beacon]

google brain yeah

[alpine coral]

yeah - mostly oai though i thought

[balmy mist]

omg 19 mins

[alpine coral]

part of deepmind's genius / plan was setup in London so as to not get all their talent lured away by silicon valley prospects

[balmy mist]

i wanted to work at deepmind so bad when I was in college 😦

[fleet lintel]

this is also sad. So much talent in EU and eventaully everything is handed over to US

[balmy mist]

it be like that bro

[alpine coral]

i really like Demis Hassabis.. was listening to this the other day driving.. worth a listen https://podcasts.apple.com/au/podcast/demis-hassabis-on-ai-game-theory-multimodality-and/id1073226719?i=1000703257125

[ember rapids]

excited for o4 mini high

[alpine coral]

he even mentions 'ultra' in passing ha

[balmy mist]

get ready bro

@keen beacon you ready?

[keen beacon]

yup 😄

[balmy mist]

im not

lol jk

[keen beacon]

smh i have to go

this always happens around oai launches

[balmy mist]

noo

[ember rapids]

gregs presenting so u know its gonna special

[keen beacon]

hopefully will be back in time

cya

[balmy mist]

what if he is a presenter?

leo's plan

[keen beacon]

my power went out during o1 preview's launch and o3 mini i think lol hopefully the trend doesnt continune

[balmy mist]

lmaoo

my youtube is not working on chrome so i gotta use brave

google dont want me to see it

[sonic tendon]

aw. cya!

i'll get on vc in 5

[plain zinc]

Feel the AGI

Guys

MMM

[balmy mist]

feelsssss

ahhhhhh

[plain zinc]

It's perfect

[balmy mist]

lmaoooo

[fleet lintel]

I feel super expensive agi

[balmy mist]

yuppp

[plain zinc]

Try FREE AGI moment

FEEL

[ember rapids]

feeling the agi

[plain zinc]

Feel his scent

[balmy mist]

remember sama said he felt the agi with 4.5

[ember rapids]

Hearing plus users get 1 o3 request per week

[plain zinc]

Lol (that's how I imagine Sam Altman)

AGI

AGI is here!

[balmy mist]

who paying $200?

[fleet lintel]

we should start protest for "free the AGI" like free the nipple movement

[balmy mist]

i am

i have to

its liek concert tickets now

damn

i stopped paying after 2 months

[plain zinc]

6 minutes

[balmy mist]

but ill do it now

[fleet lintel]

lol

[balmy mist]

feel it!!

[tawdry meteor]

how long after the announcement do we think they'll release o3 on arena so we can start ranking it against g2.5p? I want to run coding tasks across 2.5 and o3/o4mini and have them fix each other like with sonnet

[balmy mist]

yo im floatinf

omgg

[plain zinc]

PREPARE your prompts

[keen beacon]

aand i'm back :3

[balmy mist]

i thought you were gonna be in the livestream

[fleet lintel]

how many millions rich?

[keen beacon]

you got me..

i'm actually just greg's alter ego

[plain zinc]

Okay, guys. Personally, I'll check on the Google models in LMarena.

[balmy mist]

damn bro, so you getting the 20k plan?

[plain zinc]

Wish me luck (I hope I find gold)

[balmy mist]

can i share with you my prompts?

[keen beacon]

o3 and o4 mini system cards are on the cdn now

not sharing the link though 😉

[balmy mist]

ahhh

i dont see it lol

yooo

[fleet lintel]

what is the livestream link?

[balmy mist]

https://www.youtube.com/watch?v=sq8GBPUb3rk

but red is streamin it i think

[fleet lintel]

i am lubricating

[balmy mist]

i better not get blue balls i swear

[keen beacon]

lotion at the ready lmao

oh yeah

i still don't know if they're demo-ing at the end of this one

so that'll be interesting to see

decent chance that will be one last livestream this week

[balmy mist]

andwe made it

ahhhhh

[keen beacon]

sorry i forgot to say

[balmy mist]

greg!!!

[sonic tendon]

in research lounge

[keen beacon]

"demoing o4"

where's the rest of 'em

[balmy mist]

ahhhhh

[ember rapids]

https://tenor.com/view/cinema-gif-5935759059195122434

[keen beacon]

yeah okay there's gonna be another room

with

[fleet lintel]

no twink?

[keen beacon]

the agent team

[keen beacon]

stop lubricating it's over

systems 😉

he's an ai researcher what did you expect

collective slightly awkward laughs

[balmy mist]

its okay @sonic tendon , imma just watch on my laptop, cant miss this lmaoo

[keen beacon]

https://openai.com/index/introducing-o3-and-o4-mini
https://openai.com/index/thinking-with-images

[sonic tendon]

kk

[keen beacon]

lol openai's website is breaking

[balmy mist]

yeah o3 is a beast

[sonic tendon]

will prob stop streaming if you guys don't need me to

[balmy mist]

lol

[keen beacon]

just checked benchmarks vs december o3

some are better some are worse

it did worse on swe bench

by 3 points

hmph

[sonic tendon]

hmph

[balmy mist]

o3 with tools is a beast

[fleet lintel]

i want to understand the benchmark against other models... otherwise it's hard me to understand the quality

[thorny drum]

im very curious the pricing of these models

[sonic tendon]

one million dollar per token

[keen beacon]

lmao aime is basically finished now

yes but not launching today i don't think

we shall see

[torn mantle]

nah

thats crazy

they saturated the benchmark

[fleet lintel]

is this pass@1?

[torn mantle]

these tools calling are kinda neat

[sonic tendon]

peak ui design right there

[keen beacon]

It sorta reminds me of what they demod in qwq max

The tool calling

holy moly openai's website is so slow right now

either it times out or i get 500s

help 😭

[balmy mist]

lol

we gottta see vibes first

[keen beacon]

cooked

[sonic tendon]

i wonder if your secret endpoint still works

[keen beacon]

it does lol

[sonic tendon]

hey, i like it!

[torn mantle]

ChatGPT Plus, Pro, and Team users will see o3, o4-mini, and o4-mini-high in the model selector starting today, replacing o1, o3‑mini, and o3‑mini‑high. ChatGPT Enterprise and Edu users will gain access in one week. Free users can try o4-mini by selecting 'Think' in the composer before submitting their query. Rate limits across all plans remain unchanged from the prior set of models.

tf

are

you

on

????????????????????????

[keen beacon]

o4 mini on free

based

[sonic tendon]

thinking back to when leo said that o1 was 200b params

lmao

upcharge

[torn mantle]

these updates looks pretty decent

[fleet lintel]

why this feels whatever?

[sonic tendon]

i wonder if the thinking dialogue is mostly based on reinforced learning now

[balmy mist]

so do we get o3 pro?

lol

[keen beacon]

sigh.

[sonic tendon]

yeah, it happens

[ember rapids]

o3 full is cheaper than o1 i think

[keen beacon]

🙄

[torn mantle]

hes holding it so hard

quite happy guy he seems

[sonic tendon]

what an enigma

[torn mantle]

xd

[balmy mist]

hmmm

[sonic tendon]

"maximize the reasoning capabilities"
i wonder what that could mean in this context

[fleet lintel]

4x more expensive compared to 2.5 pro 😦

[sonic tendon]

local file access maybe???

[keen beacon]

brute force

BOOOOOOOOO

[sonic tendon]

?

[keen beacon]

i need to test this thing at geoguessr

[thorny drum]

4x is very good no?

[keen beacon]

he's talking about the model using the python tool to brute force an answer

[sonic tendon]

damn that's a good idea

[thorny drum]

the december version seemed to be several hundred times

[keen beacon]

https://openai.com/index/o3-o4-mini-system-card/

hey i was part of this 👀

[sonic tendon]

nice dude!

[calm sequoia]

Ahhh thats where the API is from 😄

[keen beacon]

interesting, o4 mini did best on openai's interview choice Qs

lmfao their own interview coding tasks are saturated now

[calm sequoia]

wtf

[leaden meteor]

So, openai doesn't care about arena leaderboard now? I don't see any updates today ...

[keen beacon]

i did notice that

[sonic tendon]

openai PRs? what context is this in

[calm sequoia]

In this case the o4-mini may have higher arena benchmark than the o3

[sonic tendon]

o

Measuring if and when models can automate the job of an OpenAI research engineer is a key goal
of self-improvement evaluation work. We test models on their ability to replicate pull request
contributions by OpenAI employees, which measures our progress towards this capability.
We source tasks directly from internal OpenAI pull requests. A single evaluation sample is based
on an agentic rollout. In each rollout:
1. An agent’s code environment is checked out to a pre-PR branch of an OpenAI repository
and given a prompt describing the required changes.
2. The agent, using command-line tools and Python, modifies files within the codebase.
3. The modifications are graded by a hidden unit test upon completion.
If all task-specific tests pass, the rollout is considered a success. The prompts, unit tests, and
hints are human-written.
The o3 launch candidate has the highest score on this evaluation at 44%, with o4-mini close
behind at 39%. We suspect o3-mini’s low performance is due to poor instruction following
and confusion about specifying tools in the correct format; o3 and o4-mini both have improved
instruction following and tool use. We do not run this evaluation with browsing due to security
considerations about our internal codebase leaking onto the internet. The comparison scores
above for prior models (i.e., OpenAI o1 and GPT-4o) are pulled from our prior system cards
and are for reference only. For o3-mini and later models, an infrastructure change was made to
fix incorrect grading on a minority of the dataset. We estimate this did not significantly affect
previous models (they may obtain a 1-5pp uplift).

[calm sequoia]

If o4-mini is so good, how slick is o4???

[keen beacon]

"we put in more than 10x the training compute for o1 into o3"

[balmy mist]

o4???

[keen beacon]

wtf

special surprise

agent related

here we go

yeah i presume so

yea

[balmy mist]

yupp claude code gg

[keen beacon]

anthropic gotta up their damn game

i wonder how much better this is in things like cursor and windsurf

[balmy mist]

depends on cost

the only thing holding back claude code is cost

wow

dope

[keen beacon]

lmao

what cuties

[fleet lintel]

nice demo!

[keen beacon]

"we used codex to build codex"

lol

woah

cool

[balmy mist]

anybody got codex link?

[keen beacon]

not up yet

will be in the next few weeks

[balmy mist]

i want pro now

[tall summit]

gj

[keen beacon]

i believe it's in the api now

chatgpt it is rolling out

then it's probably just a gradual rollout

higher tiers first

4.1

[balmy mist]

codex: https://github.com/openai/codex

[keen beacon]

https://x.com/sama/status/1912558745013612888

Sam Altman (@sama) on X

we expect to release o3-pro to the pro tier in a few weeks

X · Apr 16, 2025, 5:28 PM

they dont train it in

so its a hallucination

they either prompt it in the sys prompt or trained it in

waiting for karpathy's review :3

[tall summit]

coolio

[keen beacon]

they dont want u to use it 🤣

trying to make it as inconvenient as possible

[calm sequoia]

Can this thing be evaluated on arena? As I remember, tool usage is blocked

[keen beacon]

yea

i guess if u use it a lot

[keen beacon]

the regular o3 and o4 mini yeah

dk how they will implement the tool usage stuff but it shouldn't take much additional effort

let me find somethin

can u ask o4 mini this: who won the 2024 london mayoral elections and by what margin specifically? if u dont mind @deep adder

try "Let a < b < c be distinct natural numbers. Must every block of c consecutive natural numbers contain three distinct numbers whose product is a multiple of abc?"

oh

yea

cant u disable search?

go to your personalisation settings

hmm

numbers are wrong

its 1m votes for sadiq but the exact number is wrong + susan hall

expected i guess

i couldnt probe 4.1 mini for the correct numbers

is that with my prompt

yeah it's hard

my private models failed it

[tall summit]

LOL

[keen beacon]

ooh

[torn mantle]

https://x.com/DeryaTR_/status/1912558350794961168

Derya Unutmaz, MD (@DeryaTR_) on X

I’m absolutely blown away by @OpenAI’s new o3 model!

I’ve had early access and haven’t put it down for days. This release feels like the milestone we experienced with o1-preview and o1-pro, but smarter and more reliable in every way, it truly cranks everything up to eleven! In

X · Apr 16, 2025, 5:26 PM

[tall summit]

its oxygen now

[ember rapids]

Googles turn