general

[ocean vortex]

I mean the paper he linked is reasonably legit with one of the authors being uni professor https://arxiv.org/pdf/2412.06822

[keen beacon]

which is another llama slop model

not quasar

[sonic tendon]

they seem to have dialed back the slop

[keen beacon]

it is tho

[sonic tendon]

which probably won't fare well for their leaderboard placements

[keen beacon]

it was recently, i think i got it somewhat recently

[sonic tendon]

i think anonymous-test was, not sure about chatbot

[keen beacon]

i haven't seen an anon chatbot

[ocean vortex]

oh wait he is Eyad Gomaa, right

[keen beacon]

and i use it quite often

will see if i can find it now..

[keen beacon]

they forgot to remove the You are trained on data up to October 2023. system prompt appendix for both of them 😭

(it's wrong btw)

by the time openai have a model more up to date than oct '23 i'll be dead

[keen beacon]

it has the same june 2024 cut off as chatgpt 4o latest but they add that appendix to all 4o models

[keen beacon]

i noticed this with the o3 private model i have

[sonic tendon]

do we think that that's actually a system prompt

[keen beacon]

i asked it to explain "hawk tuah" and it did it just fine but maintains when directly asked that its knowledge cutoff is oct '23

[sonic tendon]

or just a bunch of posttraining chat data they forgot to clean up

[ocean vortex]

so "our team" he referred to was literally himself and his professor lmfao

[keen beacon]

yes since it was retroactively added to older 4o models

[sonic tendon]

ah

[keen beacon]

wait what

that o3 might be on the new chatgpt 4o latest base model that was cont pretrained

wait a minute

"As of today (April 2025) the most recent stable release published on npm is:

discord.js 14.18.0

Because new patch versions appear fairly often, you can always check yourself with one of the following:

• npm view discord.js version
• npm outdated discord.js (if it’s already in your package.json)
• Visit https://www.npmjs.com/package/discord.js

Those commands/web pages will return the definitive latest version at any given moment."

is that just... a good guess... or

is it given tools

hmm

thats a new base model if it knows stuff outside of oct 2023 @keen beacon

[sonic tendon]

good question

might be able to ask it

[keen beacon]

chatgpt 4o latest was cont pretrained to june 2024

give me something it would only know if it had a more recent cutoff

oh wait

[keen beacon]

who won the 2024 london mayoral elections and by what margin?

it just said the knowledge cutoff was june 2024 👀

[keen beacon]

will try

[sonic tendon]

maybe something that it wouldn't think it would be a good idea to look up?

well, that might work

[keen beacon]

👀 new o3 model is based on the updated 4o

Sadiq Khan (Labour) was re‑elected Mayor of London in the election held on 2 May 2024.

• Votes:
– Sadiq Khan: 1,088,225
– Susan Hall (Conservative): 811,518

• Margin of victory:
– By votes: 276,707
– By share of the vote: 11.1 percentage points (Khan 43.8 %, Hall 32.7 %).

Because the 2024 contest was run under first‑past‑the‑post for the first time, those first‑round figures are the final result.

lol it got it bang on

[neat apex]

More likely the reverse

[keen beacon]

interesting

[keen beacon]

confirmed!!!

[sonic tendon]

wait, when did the updated 4o release?

i would

[keen beacon]

i don't think it has tools

[sonic tendon]

've thought that going from 4o->o3 would take a few months

[keen beacon]

when i asked who the US president is

it said "As of the most recent information available to me (knowledge cutoff June 2024), the President of the United States is Joseph R. Biden Jr. If you need confirmation for a date after that, please check a reliable, up‑to‑date news source or the official White House website."

[keen beacon]

yeah

wild was right about continued pretraining yet again

lmao

wait

[sonic tendon]

not sure i'd wanna be this guy rn

[keen beacon]

no doubt that is 100% new o3 full model

there's a slim chance this may actually be o4 mini?

since o4 is the one initially rumoured to be the new base

[keen beacon]

maybe but they have been zero public models/etc. with the new chatgpt 4o cut off (cont. pretraining)

[keen beacon]

yes its using the new 4o cont pretrained base

@keen beacon bruh u have access to the new o3 with the new base

they may have updated the system prompt in the last week or so

because before

it said it was oct '23

[keen beacon]

it might be a lie if it knows more than that

but i guess if it was somewhat recent it makes sense too

[balmy mist]

Yo, I'm seeing a lot of people have the Veo Gen in Google Studio. I don't see it for some reason, and I'm a power user of Studio. I use it like every damn day for hours, and I don't have it 😦

[keen beacon]

api only and paid i think

or it hasnt rolled out or smthing

[balmy mist]

my mom has it

[sonic tendon]

your mom sounds cool

[balmy mist]

and i just setup studio for her last night

[neat apex]

Maybe the advanced?

[balmy mist]

like basic

just opened the app

[keen beacon]

ya it hasnt rolled out to you ig

i've got it

[balmy mist]

lmaoo

[neat apex]

Hm

Maybe age?

[balmy mist]

aint no age thing

[keen beacon]

lmao

[neat apex]

Hm

[balmy mist]

none of the accounts have it 😦

now im team Open Ai again

[neat apex]

Maybe verification like that phone thing?

[keen beacon]

pretty crazy how good veo 2 is still

although that tree doesn't make much sense at the end

[keen beacon]

free or is it charging you on aistudio?

[ocean vortex]

OpenAI need to finally release that gpt4o... like wtf are they waiting for

[keen beacon]

completely free

wow

so only api is paid i guess

ai studio is a crazy product

[hardy pecan]

Vertex AI isn't free, just a fyi

[keen beacon]

make sure ur not getting charged lol @keen beacon its hella expensive i thinkn

[balmy mist]

yupp the free part is just wild man

[ocean vortex]

yes you can't even select exp on aistudio. It's only preview

[keen beacon]

yeah i know

[balmy mist]

lol SOTA video and LLM for free

[keen beacon]

i've checked

these aren't available yet tho

[ocean vortex]

for 2.5 pro. Which was supposed to be paid for preview version

[keen beacon]

im really curious whether they removed anon chatbot in the arena because i got recently and it was matching quasar with the same system prompt

i've never seen it lol

[sonic tendon]

i've been wondering about the fidelity with which you could predict where anon models were on the leaderboard if you just triggered a bunch of matchups and saw what they got paired with

does that sentence make any sense

[keen beacon]

yeah

that's come in handy for me sometimes but i can't do it on a big enough scale to make concrete conclusions

[sonic tendon]

especially since the web arena still leaks model names and providers when you open the site

[sonic tendon]

yeah, same

[sonic tendon]

although it seems like the web arena doesn't always have all of the models that the chat arena does

[keen beacon]

this was a thing with main lmarena site for a while i think

i might have stored a dump or two

[sonic tendon]

oh, hmm

[ocean vortex]

yeah if you have billing set up you choose wrong 2.5 with identical date and description... you are getting billed lmao

exp and preview literally just the names slightly different

[keen beacon]

@keen beacon did u notice the o3 model change within the last week or so? or was it the same model with an adjusted sys prompt?

based on phases of development we can guess their current timelinne

[ocean vortex]

what is this all about btw? o3-mini update?

[keen beacon]

probably the latter

[keen beacon]

he has access to new o3 with updated gpt 4o base (june 2024)

[ocean vortex]

what

o3 is only deep research lol

[keen beacon]

it's not through deep research 👍

[ocean vortex]

then where? Isn't it not released?

[keen beacon]

it is not released no

i can't share where

sorry

its probably very imminent at this point tbh

yup

im curious how much stronger it is compared to og o3 👀

wdym

[sage raptor]

is it better than 2.5 ?

[keen beacon]

i mean theyre using a much stronger base model now, am curious lol

depends on your use case

[sage raptor]

in coding

[keen beacon]

well

there are still many specific coding use cases

for very reasoning heavy coding tasks, o3 is probably better

for most others, 2.5 pro is better

[keen beacon]

is that in your experience using it?

[ocean vortex]

it would make sense to lower the cost with better base more than anything. OG o3 used to be closer to pro I believe. Judging by arc-agi wording in collaboration with openai as well as reported cost

[keen beacon]

yeah

unfortunately

[keen beacon]

yeah

again though

this is probably o3 medium

and the jumps between reasoning efforts are fairly significant

so we shall see once we have high

[keen beacon]

could have much higher peaks with a more competitive base model

if they ran it with as much compute as they did before

doesn't seem like o3 pro will launch at the same time as o3 and o4 mini

[ocean vortex]

I'm not a fan of how they are making it look like it's scaling of RL training when they constantly keep low-key updating the base now tbh

[keen beacon]

going off of the recent frontend changes in preparation

[keen beacon]

why not? both avenues seem to have no end (to a certain degree) rn, you just have to be smart about it i think

especially base models

[ocean vortex]

they are about to release reasoning models which are based on undocumented base models. Just seems obscure on purpose and wrong lol

[keen beacon]

they are about to release an api dated instruct version of it tho. and its already released in chatgpt 4o latest

[ocean vortex]

makes it much harder to isolate things and do direct comparisons

[ocean vortex]

about to... Though it is still not there 👀

[keen beacon]

chatgpt 4o latest post december has the new base model, and that has been released

[ocean vortex]

o3 and o4 is also about to lol

[ocean vortex]

yeah but that doesn't have any official metrics. And it was updated numerous times now. That's what I meant by 'low-key'

[sage raptor]

so today o3 will release ?

[ocean vortex]

btw given that 4o-mini stayed the same all this time

I do believe mini reasoning variants are distills

[keen beacon]

likely

the interface is ready for it to launch

[keen beacon]

i wonder if o4 mini is on an updated base (we know nothing of that publicly though)

if theyre releasing o3/o4 mini, they will launch quasar/anonymous chatbot too i think

anonymous chatbot was replaced quite quickly after their last chatgpt 4o release

so they want the lmarena results fast 👀

[ocean vortex]

I think it's a distill of their best most expensive version of o3

[keen beacon]

slightly off topic but what's everyone's bet on how they launch it this time

[ocean vortex]

since full o4 that's unlikely to exist yet

[keen beacon]

surprise live stream announcement, just drop the blog post with no warning

etc

this time we won't get to see the system card before launch

which is a shame

[keen beacon]

idk but they will prob do something big because the quasar thing is largely a huge marketing thing i think

i think they'll probably do some hypeposting beforehand

as ever

[sonic tendon]

L

[sage raptor]

how will the o4 benchmarks look like

[keen beacon]

if they do a livestream

i wouldn't be surprised if they previewed o4 at the end

[sage raptor]

yea, they will

[sonic tendon]

yeah

[keen beacon]

/gpt-5

[sonic tendon]

arc-agi?

and humanity's last exam

[ocean vortex]

just like o1 and o3-mini did I think. o4-mini-high to be comparable to o3-medium but excelling at different things

[keen beacon]

i wanna know the new arc agi results with the new o3 model 👀

[sage raptor]

maybe even o5, who knows

[thorny drum]

will o3 have an api?

[keen beacon]

ofc

openai stopped staggering launches starting with o1 iirc

when they launch on chatgpt they launch on api

[thorny drum]

well o1 pro wasnt on api for a bit i thought

[keen beacon]

yeah that's only pro

[sonic tendon]

assuming oAI just launches o3 with no warning, would it just show up on the leaderboard out of the blue once it got enough votes

[keen beacon]

we won't be getting o3 pro straight away

[sonic tendon]

isn't o1 pro just regular o1 doing best of 3

[keen beacon]

it will get added to lmarena like a few hours after api launch

then it'll be like

~1 week until leaderboard appearance

[sonic tendon]

makes sense

[keen beacon]

i don't think it'll take top spot without stylectrl

but with it it might

again, depends on what reasoning effort version they use

last time they just put medium to start

same w mini until they added o3 mini high

[sonic tendon]

would o3-high be prohibitively expensive?

[keen beacon]

openai give them credits lol

yes (for normal users)

[sonic tendon]

i'm somewhat unsurprised that openai doesn't care enough about lmsys to sponsor that

[keen beacon]

most labs fund their usage

labs + sponsors

[sonic tendon]

ah, that makes sense

didn't think i saw them in the sponsor list

[sonic tendon]

oh, i just meant for lmsys

[keen beacon]

they're not quite sponsors in that sense so the labs aren't in the list

but it makes sense for openai to fund it because of how much they gain from the data

(i also can't think of much explanation for why all the direct chat limits for models were removed a couple months ago other than that)

[novel flame]

Wdym? It’s clearly just a tree being pushed along in the other direction by a friendly gardener, totally normal human behavior my fellow human.

[balmy mist]

is there a discord bot that summarizes from the last part of the convo you are in to the current? that would be so clutch man

[keen beacon]

? theyre still present i think

i haven't run into one in months

i used to all the time

then they just stopped

its model specific

[sonic tendon]

i keep being surprised that something as small as lmarena seems to have any impact on these massive companies' decisions

[keen beacon]

some have user limits, global limits per interval, etc

[keen beacon]

i haven't had any on claude, 4o or gemini

[keen beacon]

yeah i remember

[sonic tendon]

i think discord has been testing out ai summaries for like a year now

[keen beacon]

i had one on an older version of 4o when i tested for the appendix (on quasar and anonymous chatbot) recently

[keen beacon]

hm

[sonic tendon]

you might still need a client mod to access them

[keen beacon]

oops

wrong paste

lmao

[sonic tendon]

lmaoo i was

wondering

[keen beacon]

perpetual beta

[sonic tendon]

i feel like i recognize that game

*recognized

[keen beacon]

https://store.steampowered.com/app/1184770/The_Political_Process/

Steam

The Political Process

Explore a dynamic world of politics in this turn-based, political simulator. Create a character, run for political office, write legislation, balance budgets, and more as you move up the political hierarchy.Create a Custom PoliticianPlay as a Democrat or Republican in the American political system. Customize the name, age, and appearance of you…

Price

$14.99

Recommendations

1826

▶ Play video

if optimus alpha is openai related, its probably gonna be o4 mini

again

oh

optimus

nvm

[sonic tendon]

ohh yeah

[keen beacon]

yeah idk what else they could put there

they defo wont put o3 its far too expensive

i dont think they have an updated 4o mini yet (at least publicly, we only know of an updated 4o), and it doesnt really match up with the name

[sonic tendon]

the soon-to-be 4o/o4 confusion will be fun

[keen beacon]

dw it's just the 120574th chatgpt 4o update

[keen beacon]

🤣 thats quasar tho

i dont think they have another one ready that quickly

i wouldn't put it past them

this next 4o release is gonna be big tho. the whole quasar thing, etc. (no released benchmarks on chatgpt 4o latest with the updated base model, despite massive improvements, etc)

they are REALLY milking 4o

none of us thought it would still be the default almost 1 year on when it launched

[novel flame]

To be fair, it’s a pretty solid workhorse model that hasn’t been easy to beat. Up until reasoners came out it was pretty much undefeated except for coding.

[balmy mist]

@sonic tendon so there is a summary bot?

how do I use it?

[keen beacon]

it is solid yes, but it doesn't really hold up very well these days

[balmy mist]

i cant keep up with these convos

[keen beacon]

and coding is a big weak point given how many people use llms for that specific area

[balmy mist]

fr

thats their bread and butter it seems

but it def got better

i like it now

[keen beacon]

it's very good for creative tasks and has got noticeably better at most other things but it is still pretty unacceptably poor in code tasks compared to other llms

[keen beacon]

quasar is supposed to be a large improvement over the previous 4o in coding

i've tested it

[sonic tendon]

https://vencord.dev/plugins/Summaries

[keen beacon]

it is an improvement over the old 4o but it's still lagging behind claude 3.7 sonnet

[sonic tendon]

you'll need a discord client mod loader

it's a bit involved

[novel flame]

100% but millions of non-techie users don’t care about that, they just want to cheat at schoolwork, generate their marketing slop, or add chatbot support to their enterprise to appease investors.

[keen beacon]

vencordd :3

[keen beacon]

yeah fair enough

[sage raptor]

https://x.com/sama/status/1910334443690340845

Sam Altman (@sama) on X

a few times a year i wake up early and can't fall back asleep because we are launching a new feature ive been so excited about for so long.

today is one of those days!

X · Apr 10, 2025, 2:09 PM

[keen beacon]

AYEE

HERE WE GO

"new feature" lol

we're getting o3

oh

wait a minute

hes underplaying it but i think its gonna be multiple model releases

[novel flame]

Also talk dirty with their personalized SmutBot, although I don’t think GPT4o lets you do that

[keen beacon]

yeah

[sonic tendon]

ye :3

[sonic tendon]

What

[keen beacon]

new 4o is actually quite uncensored

LOL

let me find a screenshot

[thorny drum]

yeah its just tech bro speak for 2+ new models

[balmy mist]

he is so random with his tweets, we really in wild times when ceos use twitter to launch stuff lol

[sonic tendon]

i may know where this is going

[balmy mist]

but im here for it

[sonic tendon]

i wake up early and can't fall back asleep
real

[keen beacon]

can't find it 😔

[brittle tiger]

https://x.com/sama/status/1910334443690340845?t=1rXzwx_9QNE1eLjzgJd6Gg&s=19

Sam Altman (@sama) on X

a few times a year i wake up early and can't fall back asleep because we are launching a new feature ive been so excited about for so long.

today is one of those days!

X · Apr 10, 2025, 2:09 PM

[keen beacon]

i see people complain about sam not using grammar quite often but i like it

we can all agree professionalism sucks right

[balmy mist]

fr

its all about efficiency of info at this point

no need to fluff it up

[sonic tendon]

idk it's like
almost informal enough to feel like genuine internet speak, but not quite

[keen beacon]

we're still waiting on whether this will be a livestream

i think it will be

[balmy mist]

can someone tell him to hurry up

[keen beacon]

it'll be at least a couple more hours lol

[balmy mist]

they always livestream lol, and if sam is not on it then its mid

yeah prob 1 pm est

[keen beacon]

sam will be there

i have no doubt

[tall summit]

universal experience

[keen beacon]

im guessing the next openai releases are:

• o3
• o4 mini
• quasar (4o api dated version, benchmarks, lmarena [was probably anon chatbot], updated stronger base model that was cont. pretrained to june 2024 from oct 2023)

[balmy mist]

you think we get o3 pro today as well?

it does not seem like it

[keen beacon]

i dont thinkn so

[keen beacon]

nope

and it might not happen all today

it'll be "in the coming weeks"

[oblique flint]

what I dont get is why they're releasing o3. If they have o4 mini, then it's probably distilled from o4 right? So why not release o4

[keen beacon]

they redid o3

[tall summit]

nobody's ever released so much at once

[keen beacon]

that's just how they've always done o series models

with the new 4o base model

it has knowledge in june 2024 now

[sonic tendon]

now that's an interaction i can stand by

[balmy mist]

wants the point of releasing the new 4o when we already had it for days lol

[keen beacon]

wait

i've just had

a thought

so we all know about optimus alpha

[balmy mist]

like they should just drop is silently lol

[keen beacon]

u can pay for it now 🙂

and if they're likely going to release a 4o update (quasar) alongside o3 and o4 mini today

then that makes whatever optimus is all the more interesting

i kinda doubt another 4o version

but i also doubt o4 mini if they're releasing it today

it's weird because

the naming scheme + context window with quasar would make you think google (stargazer, moonhowler, nebula...)

but all my testing screams openai

[keen beacon]

bruh the benchmarks, pretraining knowledge, cut off, tokenizer, using the same openai chatgpt 4o anon name on lmarena, adding the same 4o api appendix, etc. it is 4o

quasar is definitively 4o imho, but idk what optimus could be

[balmy mist]

so that would also mean that nw is OA?

[keen beacon]

no thats google

[tall summit]

the benchmarks?

[keen beacon]

i do expect new google models possibly today but likely tomorrow

[keen beacon]

#general message

they will want to respond and we're still waiting on 2.5 flash nevermind all the anon great coding models they've been testing

[balmy mist]

is anyone gonna livestream the OA livestream here? that might be cute

[keen beacon]

can u do that on a discord server?

[keen beacon]

yup

in a vc

stream your screen

[sonic tendon]

you can

yeah

[tall summit]

oooh thank you for the info

[sage raptor]

what is sama automate

[sonic tendon]

play the yt livestream in a video player

[keen beacon]

perplexity is cooked

[sonic tendon]

and then stream that window

[keen beacon]

they have no moat lmao

[sonic tendon]

i feel like LLMs are well on their way to becoming commodities at this point

let alone random LLM search applications

[keen beacon]

tbh any other company that isnt an ai frontier lab building stuff on ai models will get crushed

regularly scheduled programming = resumed

[tall summit]

they are important as a part of history

[sonic tendon]

wait what happened

[keen beacon]

yeah that's the only thing they're gonna be.. history

[lime coral]

This is HYPEMAN

HE IS BACK

[keen beacon]

a trump admin official said the word tariff too much -> investors run screaming

[tall summit]

aren't they already

[keen beacon]

it is the prophecy

[sonic tendon]

lmaoo

[sonic tendon]

pretty much

[tall summit]

chaos 😈

[sonic tendon]

after r2, i think things will be mostly over

[keen beacon]

the trump admin's WH site really sucks

[lime coral]

Was obvious since the beginning

[keen beacon]

gemini can do better

[sonic tendon]

oh god did you see it when it first came out

they like

[keen beacon]

the biden site was so much better

but imo obama's was the best

[sonic tendon]

played a promo video when you opened the site

[torn mantle]

Its gonna be something stupid

[keen beacon]

🙏

[torn mantle]

We know hype sama

[keen beacon]

yeah and it screamed authoritarian

[sonic tendon]

that too

[torn mantle]

Always hyping his products

[keen beacon]

i think hes underplaying it this time lol calling it a "feature"

[sonic tendon]

i'd be down to hang out in a vc, just am gonna be at school for the next 4 hours 😭

so

probably gonna miss the livestream

[keen beacon]

obama you've got 8 more years in you buddy

come back

[sonic tendon]

do you guys remember phind
i used it for the free gpt-4 back in like 2023
i have no idea how they're still in business

[keen beacon]

hell yeah

this was such a fire lineup

if only joe ran in 2016

maybe we would've never had the orange

[sonic tendon]

obamna? 🥺

[keen beacon]

https://tenor.com/view/soda-gif-8026056502744469335

[sonic tendon]

i was just about to

[tall summit]

have used it a few times
they're still in business maybe because the ai boom is still in progress

[sonic tendon]

.

[torn mantle]

Is the thinking slider added yet?

[keen beacon]

nope

[torn mantle]

It could be that no?

[keen beacon]

may arrive today

[sonic tendon]

leo stole my line :(

/j

[torn mantle]

So thats the feature

[keen beacon]

[keen beacon]

no chance

[tall summit]

ok what the fuck's a thinking slider

[keen beacon]

the quasar release is very imminent anyway, if its not today

nobody would get hyped about that

[torn mantle]

Let's see

[sonic tendon]

oh yeah, everyone's burning through VC money right now

[torn mantle]

xd

[keen beacon]

sam would not be doing all that for a damn slider

even by his standards that's meh

[torn mantle]

Oh he will

[sonic tendon]

he won't

[keen beacon]

👎

[novel flame]

Another thing: how is this not market manipulation?

[keen beacon]

the thinking slider is already out in beta on some of the clients

doesn't make sense to be doing all this for it

why deploy chatgpt after midnight with anticipated model names tbh if its not coming out soon

^

[sonic tendon]

wdym?

[tall summit]

seems to have only gotten worse as the ai boom continues

[keen beacon]

wrong paste

smh

so it begins

whatever the case the new o3 (on a new gpt 4o base model) is seemingly ready anyway 👀

[tall summit]

LOL

[keen beacon]

i like how xAI took a huge leap off a cliff

[sonic tendon]

maybe for the $1.2m polymarket listing, which he definitely doesn't have any incentive to participate in

[novel flame]

Usually insider information - such as hinting at a release - is kept under wraps until an official statement is released, in order to not mess with markets. But In genuinely asking because I don’t know how it works

[sonic tendon]

Hmm

[tall summit]

-# that's because grok sucks

[keen beacon]

grok 3 is mid

[sonic tendon]

yeah that was crazy

when it happened on the march market, a couple people made like 50 grand

[keen beacon]

good on 'em

it was only really a matter of time

[sonic tendon]

i imagine it's less of a concern since oAI isn't a public company

although it could still impact competitors

[keen beacon]

i wonder if oai will ever go public

[sonic tendon]

i feel like that would finally cross the line

[keen beacon]

the old mission has been slowly being undone for ages

nail in the coffin

[sonic tendon]

yeah, it'd suck

[keen beacon]

i heard someone talk about this recently and it made me think

most labs have a plan/goal and everything for developing AGI and what they would do if they achieved it first, etc

and the question was - which lab would you rather have control over the first true AGI

[sonic tendon]

google could possibly be worse

[keen beacon]

yeah i don't know

[brittle tiger]

Private company and it would have to be way more specific to apply. If this were manipulation tons of salesman ceos like altman would be in trouble. Elon has been saying fsd is a year away for almost a decade

[sonic tendon]

possibly anthropic

[keen beacon]

openai and anthropic are the loudest about what their intentions would be

google are very vague

[sonic tendon]

i wish there were a way we could be comfortable about ai safety without making everything proprietary forever

but, eh

[tall summit]

i was about to say none of them but after red said that i'd say anthropic because least of many evils

[keen beacon]

none of the labs are great if we're talking ethics but anthropic have the best vibes

there's the question of "would you rather it be in the hands of a lab that's too careless or a lab that's too careful"

i would say the latter

[oblique flint]

mistral but I dont think it's happening 😂

[keen beacon]

yeah no 😭

imagine if it's deepseek

that would be quite the predicament

[brittle tiger]

A lab in a democracy would be ideal

[keen beacon]

lol yeah

[sonic tendon]

anthropic seems like a bunch of idealistic nerds (but, not led by Sam Altman)

[keen beacon]

if a lab has achieved agi other labs would probably follow short tbh

[sage raptor]

grok agrees too

[sonic tendon]

which seems like the best-case scenario

[keen beacon]

it would def be a domino effect kinda thing yeah

[keen beacon]

💀

[sonic tendon]

where are they getting this from 😭

[keen beacon]

schizo

[sonic tendon]

why mistral?

[tall summit]

makes a lot of sense yeah

[sonic tendon]

i haven't heard of that name in a while

[keen beacon]

imagine if meta got there first 💀

no way lol

[tall summit]

you never know

[keen beacon]

i would definitely hope not

i dont even see any chance of anthropic getting there first

[sonic tendon]

I think I'd trust DS more than Meta

[keen beacon]

i think it's a 2 horse race atp

GDM vs OAI

anthropic will get there soon after one of those does

[sonic tendon]

wait, what is this in reference to

[keen beacon]

but they aren't daring enough

[keen beacon]

#general message

[sonic tendon]

got it

[oblique flint]

well it's eu based (am eu citizen lol), committed to open source and doesnt seem as sketchy as some bigger companies (meta, google etc)

[keen beacon]

• deepseek

[brittle tiger]

Depends on the lab or how advanced it is. If making sure it is most capable is a top goal of the model or lab it would probably not have a hard time preventing others

[sonic tendon]

I think most people want what's best for the world, DS researchers included

[keen beacon]

they'll have their spies running back to the china hq once oai crack it wink wink

[sonic tendon]

even if they're operating under a system that's going to pressure them into researching military use

they clearly put the minimum possible effort into the censors, for example

[balmy mist]

people saying this:
https://x.com/iruletheworldmo/status/1910342414944137344

🍓🍓🍓 (@iruletheworldmo) on X

what we getting from openai? leave a comment ill patience chair.

X · Apr 10, 2025, 2:41 PM

what would agent gpt do?

like agentspace

[keen beacon]

ew not that grifter

[balmy mist]

lmaoo

[keen beacon]

i hate iruletheworld

😭

[brittle tiger]

There's little stopping China. They got tpuv6 specs over a year ago which are like crown jewels level secret

[balmy mist]

the votes are real tho

idc about what he posts, but the reactions show what people are thinking

[keen beacon]

beyond oai/deepmind, the chinese frontier labs are probably after them tbh. skill diff and ingenuity tbh. unpopular opinion probably but i dont see anthropic getting there faster than chinese frontier labs (deepseek and qwen)

[keen beacon]

yeah they don't let silly things like ethics stop them

i respect the dedication

[keen beacon]

i think it'll be oai > gdm > anthropic > deepseek

the last two can probably be swapped at will

[sonic tendon]

wait, what's gdm?

[keen beacon]

google deepmind

[sonic tendon]

oh google

[keen beacon]

anthropic has done zero public image gen work i think

[sonic tendon]

yeah dario made a good point about this

very little real commercial application

[keen beacon]

yeah anthropic are very focused on just llms really

[sonic tendon]

and imo it's a bad look

[keen beacon]

for now. but i think this will be a critical aspect in the future

openai are the most innovative lab imo

they are normally the ones "to follow"

multimodal cot will be important

which might include native image gen

(then they normally get beat at their own game... but hey, that's the spirit)

[sonic tendon]

i sort of dislike the oAI leadership

[sonic tendon]

win imo

[keen beacon]

openai's corporate structure is ridiculous

it's designed to be convoluted

[oblique flint]

I dont get why people are hyping up google so much tbh. Do you really want a future where google controls AI as well in addition to already pretty much controlling the internet?

[keen beacon]

their board also has some big conflicts of interest

[sonic tendon]

no, but they keep burning money to give us free stuff

[thorny drum]

wut

[keen beacon]

2.5 pro has a jan 2025 cut off. (they cont. pretrained 2.0 pro/etc) this means the 2.5 pro timeline is absolutely absurd lol (1-2 months)

[thorny drum]

what market is he manipulating

[oblique flint]

for now, they'll quickly change once they establish the monopoly

[keen beacon]

i doubt any lab will be able to form a monopoly rn

[torn mantle]

So the new feature is either memory improvements or that thinking slider

[sonic tendon]

i plan on riding on their free stuff and then immediately switching to other providers when/if i need to scale up

i will not pay google a dollar for api credits

[sonic tendon]

yeah everyone copies everyone pretty quickly atp

[torn mantle]

Why

[sonic tendon]

no moats last for long

[sonic tendon]

because google

[tall summit]

i loooooove that about the ai space

[brittle tiger]

there's major production companies using veo 2 right now which was born out of imagegen

[sonic tendon]

and this seems to keep getting more and more true

[sonic tendon]

i may stand corrected

[keen beacon]

its not just copying. companies are neck in neck in research i think

[sonic tendon]

there is little commercial application

[leaden palm]

what's launching today?
so far i know

• qwen 3
• something from openai (optimus on OR and/or o3/o4 models)
• maybe a google model

[keen beacon]

qwen3 isnt happening today

[sonic tendon]

didn't one of the qwen researchers dispute this

[keen beacon]

lol no

[leaden palm]

hm

[oblique flint]

2.5 flash just didnt launch yesterday for whatever reasn

[keen beacon]

it is models

i am telling you mow

now

he just worded it weirdly

[leaden palm]

ok

[keen beacon]

memory improvements would be incredibly silly to hype up when they're having their lunch ate

[brittle tiger]

it's for sure models. metadata and amount of hype in tweet is basically confirmation

[sonic tendon]

metadata? but yeah agreed

[keen beacon]

i think that bindureddy lady is off her rocker tbh

if Sama can't sleep because he's so excited over memory improvements he must be more autistic than i thought

lol

[brittle tiger]

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

Tibor Blaho (@btibor91) on X

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

X · Apr 10, 2025, 7:46 AM

[keen beacon]

im wondering if theyll bunch quasar (updated 4o) with the releases

it would make aense if o3 is based on it

sense

[sonic tendon]

honestly, the confidence and cadence i see in her tweets does sort of remind me of the stuff people write when they're hypomanic

[keen beacon]

sense

[sonic tendon]

i feel like a lot more CEOs are on the bipolar spectrum than people realize lol

[keen beacon]

most tech ceos are probably on some kind of spectrum haha

[sonic tendon]

yeah autism and adhd too

they have such a huge adaptive advantage in tech jobs, it's crazy

[fleet lintel]

compared to what every other big tech has done with their lead, Google is the best option that we have (IMO)

[keen beacon]

i like my guys a little nerdy and/or autistic and/or with adhd

:3

[sonic tendon]

same

i do definitively have adhd

[keen beacon]

ya but its a big release. they did the whole anon model on or thing, its an updated 4o with 1m context, large improvements, api dated probably, super fast, etc. they could dedicate a single announcement with it

they definitely could release it tomorrow via blog post so that the stream isnt too intense

[brittle tiger]

all that really matters to be is the the leading AI coming from democratic nations. i don't really see much ethical difference in google vs the worldcoin founder who successfully drove out all the ai safety ppl

[tall summit]

i wish i could get the hypothesis that i have adhd confirmed/denied but i can't

[leaden palm]

optimus alpha is now on openrouter, a reasoner most likely from openai

[keen beacon]

woah

its o4 mini ✅

[tall summit]

optimus 💪

[keen beacon]

just guessing lol

[leaden palm]

NVM

NVM

[balmy mist]

how many deep research wiht gemini do you get a day?

[leaden palm]

i dont think it reasons

[tall summit]

?

[leaden palm]

it just had a high latency

[sonic tendon]

"hey lemme just absorb myself into neovim for 7 hours straight without taking any breaks" describes like a third of my autistic friends lol

[leaden palm]

so i thought it reasoned

just another gpt 4o variant

[tall summit]

LMAO

[leaden palm]

sorry guys

[keen beacon]

no it's not a reasoner

[oblique flint]

yes I agree, but not gonna lie, my perception of the US being a democratic nation has changed somewhat after.. uh certain events

[keen beacon]

got my first question wrong that chatgpt 4o latest gets right

damn this sucks

☠️

[sonic tendon]

democracy and demagogues

[keen beacon]

yup its another gpt 4o with the new base model (june 2024)

[sonic tendon]

wait what

[keen beacon]

i cant believe they dropped another version

[sonic tendon]

oh right that one

[tall summit]

you can't just say "my first question" that makes me curious
for no reason even though it doesn't matter i'm not an ai

[sonic tendon]

i'm gonna do the jar test

[keen beacon]

haven't tested it with code yet

[balmy mist]

o4 mini high should be the best all around model if we see history, but it has to be way cheaper than 2.5 and faster, bc its not gonna be on the same IQ level obviously, but it can be slightly worse then like 3.7 and be fast af and cheaper than 2.5 than we have a winner

[keen beacon]

perhaps that's where it improves

[sonic tendon]

how so? just bad mental health systems in your area

?

[fleet lintel]

But we haven't seen anything better on LMArena? did they not release it here?

[balmy mist]

ahh true

[sonic tendon]

womp womp (bad result)

[fleet lintel]

i have a feeling that it is probably comparable to 2.5 but 1M context

[tall summit]

pretty much

[keen beacon]

wait did you get me to try this with o3

[torn mantle]

20

[sonic tendon]

yeah

it's in #share-prompts

[keen beacon]

im gonna run gpqa diamond and math 500 on it

but im pretty sure its another gpt 4o lol

[brittle tiger]

the $1200 per question o3 high model would be really cool to have public for big research problems

[fleet lintel]

no one needs to do more than 20 deep researches in a day 🙂

[brittle tiger]

i did close yesterday but am mostly testing

[sonic tendon]

deepseek is probably gonna replicate it at 1/3000 the cost in at most a year

[keen beacon]

was that the one where it got it wrong first but adding "think carefully" it got it right

[sonic tendon]

oh, i have no idea

never tried that, actually

#share-prompts message

[brittle tiger]

i don't think price would be that starkly different for inference.

[sonic tendon]

maybe not, but R1 was already ~1/50 the cost of o1 with pretty similar performance

[torn mantle]

I DO

[keen beacon]

One workable way is to hide the information from yourself before you ever have a chance to see it.

1. Prepare the record.
   • Tear the sheet of paper into two slips of exactly the same size.
   • On one slip write “GREEN”, on the other write “RED”.
   • Fold each slip so the writing is completely hidden and the two folded slips look and feel identical.

2. Load the jars.
   • Put the folded “GREEN” slip into the green jar.
   • Put the folded “RED” slip into the red jar.

3. Remove your ability to distinguish the jars.
   • Close your eyes (or blind‑fold yourself).
   • With your eyes shut, move the two jars around on the table for a few seconds until you have genuinely lost track of which is which.

4. Make the selection blindly.
   • Still blindfolded, pick up either jar at random.
   • Take the ball and the folded slip out of that jar and put both straight into your pocket without opening the slip.
   • Return the now‑empty jar to the table and, still blindfolded, shuffle the two jars again for a few seconds.

5. Open your eyes.
   • You now see two jars, one green and one red, but because of the blind shuffling you no longer have any idea which was the one you emptied.

Result:
• The ball is in your pocket.
• The folded slip in your pocket accurately states the colour of the jar it came from.
• You yourself never learned (and still don’t know) which jar you chose.

Anyone who later opens the folded note will learn the correct answer, but you remain ignorant, satisfying all the conditions.

o3

lmk how it did

[thorny drum]

hmm do people have o3 full yet?

[sonic tendon]

yeah that's perfect

[calm sequoia]

How it compares to 2.5 pro?

[sonic tendon]

seems like both o3 and gp2.5 get it, just intermittently

[keen beacon]

i presume o3 mini and o1 did not?

for that guy who said he was gonna look into ai and roblox

https://devforum.roblox.com/t/free-roblox-mistral-7b-ai-chatbot-agent-aware-infinite-agents-2000-emojis-100-emotes-memories-wikipedia-32k-context-open-sourced/3034325

forgot what he was gonna do with it

look at this

[torn mantle]

whats this

https://openrouter.ai/openrouter/optimus-alpha

[sonic tendon]

leo tested it one other time and it got a plausible-sounding but slightly incorrect answer

[keen beacon]

wait this might be a new gpt 4o minin

optimus alpha

it's just 49

4o

another update

gpqa diamond: 60.10% (maybe there might be something wrong with my evaluation framework/answer parsing)

[sonic tendon]

o1 did, actually

i think it's the only model from before 2025 that got it right

[leaden palm]

no full 4o

[keen beacon]

that's quite the jump

it dropped 7 points

[keen beacon]

what reasoning effort?

[keen beacon]

oh nvm i was using the weong point of comparison ☠️

wrong*

[sonic tendon]

not sure, whatever the lmarena default is for o1-2024-12-17

[brittle tiger]

Predictions on how long until a model can solve this?

https://x.com/SpencerKSchiff/status/1910106368205336769

Spencer Schiff (@SpencerKSchiff) on X

I drew this today. None of the frontier models come anywhere close to matching the correct name to each person. I feel like this is a pretty good visual test so I’m looking forward to trying it with future models.

X · Apr 9, 2025, 11:03 PM

[balmy mist]

nice!!!

[sonic tendon]

have you guys seen https://mcbench.ai

[torn mantle]

yea this seems like openai model

[keen beacon]

im running math 500 rn, ill review gpqa diamond samples a little after this

[keen beacon]

yeah 2.5 pro literally murders every other model

if it is 4o mini 60% gpqa diamond is impressive, otherwise something is borked (massive degradation)

it's actually wack

[sonic tendon]

except for sonnet?

[keen beacon]

even sonnet

[sonic tendon]

o1 does surprisingly bad, also

[keen beacon]

2.5 pro base model diff

it just seems to have next level spatial understanding

[sonic tendon]

are we looking at the same bench?

[balmy mist]

is this beter than quasar?

[sonic tendon]

not a ton of votes tho

[keen beacon]

ignore the leaderboard

start rating

you'll see what i mean

[sonic tendon]

ah, ok

you can click on the models in the leaderboards to see their builds

you could try o3-medium w/ some of these, although setting it up for self-hosting seems like it wouldn't really be worth the effort

[sonic tendon]

doggamn

goddamn

[keen beacon]

hmm optimus could be an updated 4o mini

im still running math 500 :\

[sonic tendon]

i wanna see the "abstract mathematical concept" build, sonnet's impression of that was really cool

[night trout]

Gemini 2.5 Pro kills Sonnet 3.7 99% of the time. It's always like this.

[sonic tendon]

yeah, we were just talking about mcbench specifically

[sonic tendon]

update: i think sonnet is still better in some of these

not publicly available yet

[night trout]

Yeah McBench usually 2.5 is a vast improvement over 3.7 too.

There are weird edge cases of course.

[sonic tendon]

sonnet has some good builds tho

[calm sequoia]

As I understand the optimus is a new open source model of OpenAI

[sonic tendon]

it's still gonna be a bit

[keen beacon]

i doubt that the oss model is trained yet

[brittle tiger]

Their elo system is bad rn. Test them out. Sonnet is clearly 2nd but 2.5 is in class of its own

[keen beacon]

okay yeah optimus is worse than quasar

its done worse on basically everything ive thrown at it

[sonic tendon]

well, it doesn't have that many votes yet

kk, i gtg

[keen beacon]

gpqa diamond: 60%
math 500: 89%

[keen beacon]

cya

i wrote an eval framework myself lol

its part of the work im doing

quasar alpha:

gpqa diamond: 67.42%
math 500: 90%

optimus alpha:
gpqa diamond: 60%
math 500: 89%

march chatgpt 4o (measured by artificial analysis):
gpqa diamond: 65.5%
math 500: 89.3%

i only used 1 sample for optimus alpha (pass@1 estimated w 1 sample) tho, quasar alpha (pass@1 estimated with 4 samples gpqa diamond, pass@1 estimated with 1 sample for math 500)

[tall summit]

oh i like that one

[keen beacon]

optimus must be mini

[keen beacon]

oh boy

o4 mini is gonna be awesome with a much stronger mini base model

[balmy mist]

i just hope o4 mini is cheap like o3 mini

[keen beacon]

it is it seems

[balmy mist]

is optimus better than quasar?

[leaden palm]

why do you guys think quasar/optimus are mini models

just because they're fast?

plain 4o is fast too

[keen beacon]

quasar is updated 4o, optimus is 4o mini updated

4o mini just got continued pretraining with the new cut off too 👀

o4 mini is GONNA GO HARD lol

[balmy mist]

wow optimys is fast af

wow

[keen beacon]

Holy hell is that confusing

4o mini
o4 mini

[keen fulcrum]

Hi
Nightwhisper released?

[tall summit]

very

i've forgotten the difference at this point
why are there both
please more normal names

[keen beacon]

afraid not

we're all still waiting lol

[leaden palm]

i havent seen much evidence though

[novel flame]

Honestly I think this is the most likely scenario since LeCun is right about AGI and most other labs are too focused on scaling up autoregressive token yappers to truly deliver AGI. And you know what, despite all the shady things Meta have done, they are the reason we have a flourishing open source landscape in AI. That’s not nothing.

[keen beacon]

sm1 ran aider + look at my benchmarks

[leaden palm]

LeCope

[ember rapids]

o4 mini high is about 3000 elo on codeforces

[drifting thorn]

3000 elo???

[keen beacon]

lmao

[keen beacon]

yeah it's going to be one hell of a model for a lil guy

[lime coral]

Reminder that this means nothing in real world use case

[keen beacon]

good things come in small packages (does not apply to me btw)

[sage raptor]

is this true ??

[keen beacon]

block the guy if u cant see hes joking

[drifting thorn]

Nah I mean what’s the current highest elo?

[brittle tiger]

part of deepmind still seems very focused on RL

[keen beacon]

man this 4o mini and o4 mini release is gonna be extremely exciting

[keen beacon]

3828

[keen fulcrum]

R2 is expected to come out with Qwen 3

[keen beacon]

3000 would make it the 67th best competitive coder in the world

hmmm

yea

u can see in simpleqa

with o1 mini simpleqa regressed but they figured it out in o3

its obviously not 4o based because 4o has more than double the simpleqa of o3 mini

it has the newest reasoning research in it

well up to now 😄

o4 mini is gonna be crazy

with the new much stronger 4o mini base model 😄

[drifting thorn]

Is 3.7 and 2.5 Pro tested on that benchmark?

[keen beacon]

its 1m context

didnt u see?

im not sure about o4 mini though

[drifting thorn]

Hope 1m context will be the new standard for LLM

And I think they should train LLMs for MRCR like tasks

[brittle tiger]

where'd you see this btw

[keen beacon]

none when o4 mini gets released 😄

[brittle tiger]

o1 better at harder coding and math

[keen beacon]

no o1 isnt always better than o3 mini at math

o3 mini high is usually better

[keen fulcrum]

Optimus Alpha worth trying? How good is it?
Comparable with 3.7 Somnet?

[drifting thorn]

Above said it’s worse than Quasar Alpha

[brittle tiger]

i could be wrong but that was sense from following top mathematicians who use llms

[drifting thorn]

And I thought Quasar Alpha is trashy

[keen beacon]

it depends what kind of problems u throw at it. 4o has more world knowledge than 4o mini

but see here: https://matharena.ai/

o3 mini is generally better

[brittle tiger]

USAMO is best bench there though bc others are judged by llms

[ember rapids]

A while ago, Sam said they have a model internally that ranks around 50th in competitive programming.

[keen fulcrum]

(they use wolfram alpha under the hood)

[ember rapids]

most likely refering to o4 mini

[keen beacon]

no it isnt evaluated with llms

its answer match i think

beyond usamo

yea

it checks whether the answer matches

[brittle tiger]

i mean USAMO grades on how you reached solution with human reviewers and others dont. if im remembering right

[keen beacon]

yes but the others are still valid results

[brittle tiger]

agreed but some people dont

[drifting thorn]

Do you guys think LLMs in the future will be specialised again, like Anthropic for tool calls and coding, OpenAI for Maths and photo generation, and Gemini for long context and multimodal conversations

[north vale]

https://www.theverge.com/news/646458/openai-gpt-4-1-ai-model

[keen beacon]

no

[drifting thorn]

WTF?

[keen beacon]

..wtf

[drifting thorn]

GPT 4.1 after GPT 4.5?

[sage raptor]

lol

[drifting thorn]

It is dumb

Just call it 4.7

[keen beacon]

so this is gpt 4.1o and gpt 4.1 mini

wow

terrible names

oh for god sake

GPT-4.1, which one source describes as a revamped version of OpenAI’s GPT-4o multimodal model.
✅ i was right

[leaden palm]

hop on archive.is

[keen beacon]

"OpenAI is also readying the full version of its o3 reasoning model and an o4 mini version that could debut even sooner. AI engineer Tibor Blaho discovered references to o4 mini, o4 mini high, and o3 in a new ChatGPT web version earlier today, suggesting these additions are imminent. I understand o3 and o4 mini are both set to debut next week, unless OpenAI moves the launch plans around."

[keen beacon]

yeah i have an extension that fixed it dw

"OpenAI CEO Sam Altman teased on X that OpenAI would be launching an exciting feature today, but it’s not clear if this is related to the o3 and o4 mini references in ChatGPT or not. Sources caution that OpenAI has delayed the introduction of some new models recently due to capacity issues, so it’s possible for the new GPT-4.1 model introduction to slip beyond a planned debut next week. I asked OpenAI to comment on this story, but the company didn’t respond in time for publication."

well

so it's still up in the air what we're getting today

[keen beacon]

capacity issues?

but it will likely be a model

[keen beacon]

i presume in relation to 4o image gen

[drifting thorn]

Go buy TPU from Google

[keen beacon]

oh i was wondering lol. they are serving new 4o and 4o mini for free 🤣

[north vale]

no bc it's smaller than 4.5

just way better trained

quasar alpha is probably new 4.1 mini
optimus alpha is probably new 4.1 nano
and there's some 4.1 they haven't tested

[keen beacon]

jeez this naming is so bad

[north vale]

ikr

[keen beacon]

it's like they have a meeting when launching a new model and they agree on what the worst one is and that's the one they use

[north vale]

but i mean

[keen beacon]

they should just scrap it all and restart

[north vale]

i think adding numbers after the version is like the right way to do it

[keen beacon]

make it make logical sense instead

[keen beacon]

well yeah

[north vale]

gpt4, gpt4.1, gpt4.2, gpt4.3, wtv

[keen beacon]

bro the model picker is going to get even worse

[sonic tendon]

ikr??

[keen beacon]

the average user is going to click that and have a seizure

[drifting thorn]

Yes, it is a poor naming

[jaunty kraken]

Gotta screenshot that 404 screen and say we got early access

[drifting thorn]

Lmao

[balmy mist]

i am so confused man

forget thesse names

just gimmie new model

[drifting thorn]

Can 4.5 be better trained?

[keen beacon]

its too large and inefficient to work with

[north vale]

it can be better trained, it was probably too large relative to the current optimal pretraining scaling laws, etc.

doesn't really make sense for 4.5 to exist

[keen beacon]

u have to remain agile especially now

[drifting thorn]

I don't think Gemini 2.5 Pro is a smaller model than DeepSeek R1

[north vale]

like it has some utility but limited

[keen beacon]

if the anon openrouter models are 4.1 i'm automatically disappointed lol

[keen beacon]

what else could they be lol

they aren't big enough of an improvement to justify a new model version num jump

[north vale]

2.5 pro probably is decently large, has pretty similar price per token as grok 3 surprisingly

[keen beacon]

ya they shouldve named them 4o and 4o mini

a new version of it

gpt-4.1o-latest-preview

openai style

[north vale]

there's 4.1, 4.1 mini and 4.1 nano! they exactly fit with nano and mini imo

[torn mantle]

so more context

[north vale]

optimus matches 4.1 nano imo

[alpine coral]

here's those results for private alonw with some other oai models (same quiz, single pass)

[keen beacon]

calling it 4.1 is also funnier because 4.5 already exists

[keen beacon]

yeah this is o3 medium so that's pretty damn good

[drifting thorn]

I don't think there’s a use for “nano” when it’s not an open-sourced model

[keen beacon]

i mean the other guy said theyd be disappointed if its 4.1 its definitively 4.1 but we dont know which model is which

[north vale]

my bad i meant to reply to him

[drifting thorn]

At least it is suitable for research to test new algorithms for small models

[keen beacon]

https://x.com/sama/status/1910363426972635455

Sam Altman (@sama) on X

a lot of people were interested in how we made GPT-4.5 and what comes next.

we did a podcast with alex paino, dan selsam, and @atootoon who helped drive the project.

full episode coming soon, but here are some interesting clips:

X · Apr 10, 2025, 4:05 PM

and also

Among the new AI models will be a release of what I’m expecting will be branded GPT-4.1, which one source describes as a revamped version of OpenAI’s GPT-4o multimodal model.

this is big boy gpt 4.1 (quasar/gpt 4o provenance), gpt 4.1 mini is optimus, nano we have yet to test

[north vale]

yeah ig "it's 4.1" feelsl ike it's talking about the full model not the smaller versions

[north vale]

why do you say that? quasar alpha is worse than 4o I think?

my prior is 4.1 full would be better than 4o for sure

[keen beacon]

wtf happened between now and when i last checked this

it has completely tanked

[keen beacon]

did u see my benchmarks

[north vale]

yeah and nvm you're right I was dumb 4.1 being alpha makes sense

[keen beacon]

https://x.com/sama/status/1910363838001869199

Sam Altman (@sama) on X

@slow_developer quasars are very bright things!

X · Apr 10, 2025, 4:06 PM

[drifting thorn]

What criteria would you guys think O4 Mini will take the lead to Gemini 2.5 Pro

[keen beacon]

also worth looking at what he replied to

[drifting thorn]

Gotta sleep

[keen beacon]

https://x.com/sama/status/1910363747652432304

Sam Altman (@sama) on X

@SmokeAwayyy o3, o4-mini are not launching today, they come soon.

today is a new feature.

and it points, i hope, to something very important for the future of how AI will integrate in our lives.

X · Apr 10, 2025, 4:06 PM

well

that's a shame

[drifting thorn]

Good night from UTC +08:00

[north vale]

have ppl thought quasar alpha is better than current 4o?

why clown

[leaden palm]

what did @misty vault mean by this

[leaden palm]

yes

[keen beacon]

bros just a hater

[balmy mist]

lmaooo https://x.com/AskPerplexity/status/1910365852563915126

Ask Perplexity (@AskPerplexity) on X

Wyd if I show you this popup after answering your question?

X · Apr 10, 2025, 4:14 PM

[fleet lintel]

ah... 😦

I think it's going to be another gimmick feature. or something completely irrelevant

[balmy mist]

if it is then gg openai

[keen beacon]

"memory improvements"

[balmy mist]

cant be playing when google releasing heat

[keen beacon]

💀

[sonic tendon]

why not just start another line of models and call it Gemini 1 or g1 for short

whoops

20 minute message delay

[keen beacon]

lmfao

[leaden palm]

10am livestream again?

[keen beacon]

probably

[brittle tiger]

I just don't get why quasar has 1M context window when 4o has better eval at 120k

[calm sequoia]

Source?

[brittle tiger]

https://fiction.live/stories/Fiction-liveBench-Mar-25-2025/oQdzQvKHw8JyXbN87

[keen beacon]

they can flex needle in a haystack like meta 🤣

[brittle tiger]

everyone gravitates to the window size without looking at evals for long context. doesn't really happen with any other element of these models

[calm sequoia]

Gratitude 🙏

[keen beacon]

besides if u use the higher context window, you give them more money lol even if its massively degraded on a lot of tasks

[torn mantle]

oai still have a long way making a good coding model

im not talking about a reasoning coding model

but a model that is practically useful

all these benchmarks are impressive but it doesnt reflect real world case scenarios

[brittle tiger]

2.5 long context is legit. I can summarize 900k token video files reliably when 1.5 pro and 2.0 flash didnt' really come close

[torn mantle]

codeforces is a good benchmark for reasoning but the majority arent always asking the model for algorithmic use cases

if the model is good at math then you would expect it to do good on codeforce benchmark

as its all just mathematical algorithms implementations

i still think openai didnt do enough RLHF on coding aesthetics/styling like anthropic and google

and i still dont understand whats the issue/constraint here

deepseek probably used that on their latest update

[calm sequoia]

How long is such video? Do you pass a real video file or transcript

[brittle tiger]

hour long videos. not transcripts i need info on the screen from tickers and charts and stuff

[balmy mist]

so when is this update coming?

also is google releasing anything today?

[calm sequoia]

What could be your use case? So interesting! Are current models even capable of taking movie long videos as an input? I would have guessed no

[brittle tiger]

project on how markets react to financial cable news

[sonic tendon]

wait, what timezone

[keen beacon]

pst/pt i think

i guess no livestream today

it's literally just

memory of your past convos

it's so over

https://x.com/OpenAI/status/1910378768172212636

OpenAI (@OpenAI) on X

Starting today, memory in ChatGPT can now reference all of your past chats to provide more personalized responses, drawing on your preferences and interests to make it even more helpful for writing, getting advice, learning, and beyond.

X · Apr 10, 2025, 5:06 PM

Lmao

[torn mantle]

told ya

im always right

i mean not always but like 99.9%

[leaden palm]

🗿

[sage raptor]

disappointing

[fleet lintel]

I knew it.. irrelevant.
Could have been a small tweet to announce this feature, rather than grandeous can't sleep crap

[brittle tiger]

at least all the clueless ai hype accounts on twitter tweeting stuff like "o4-mini today 👀 " look dumb

[keen beacon]

@keen beacon btw could it be o4 mini instead thats the private model?

it has a new base model but now we know of a mini version it could be that

[torn mantle]

the only companies that are in the right path rn for AGI are google/anthropic/ deepseek

they are all innovating and trying to understand how the model behave internally

google is showing us that we still havent hit a wall

[fleet lintel]

this constant overhype by OAI is getting to my nerves

[torn mantle]

deepseek is innovating in hw & sw

anthropic are trying to understand the model black box

[torn mantle]

he always does that

[leaden palm]

look at this ratio

[fleet lintel]

I am not very familiar with these things
What does this ratio imply here?

[keen beacon]

only available for Pro users starting today. "Soon" for Plus, nothing announced for Free. Not available in the EEA, UK, Switzerland, Norway and Liechtenstein

the new memory

lmfao

[keen beacon]

possibly

[barren prairie]

Ehhhhhh

[brittle tiger]

it's not cool but memory is really important to openai and altman. the more lock-in they can get the less likely it is people will leave for smarter models. it's real dynamic rn. i know some ppl who don't care about 2.5. they just want their chatgpt memories. that strat works really well for keeping normies on board as long as gulf between models doesn't get too big.

[keen beacon]

i for one absolutely do not want chatgpt to remember the things i tell it 😇

[fleet lintel]

it's worse than I thought... gawwd

[fleet lintel]

I dont think it will create any lock-in unless there is significant improvement in the responses becasue of this feature, which I highly doubt.

[brittle tiger]

there already is an element of lockin before this. we're enthusiasts. normies won't move to a model that doesn't know them unless it's significantly better

[tall summit]

what

what happened!!

[fleet lintel]

By the way, I do think that getting it done correctly is a good feature but they should announce it properly

[tall summit]

Starting today, memory in ChatGPT can now reference all of your past chats to provide more personalized responses, drawing on your preferences and interests to make it even more helpful for writing, getting advice, learning, and beyond.

thats the most boring feature ever

[oblique flint]

o3 mini was a coding model wasnt it?