general

[alpine coral]

the rope / crabiner one - i need to refine that

yeah no doubt all models in theory would do better at each question individually

though part of the 'test' itself is the ability to answer all the questions at once

[and in the case of the last 'question', refrain from generating a list.. which is the only critereon on which the response is judged ha.. so many models strugggle]

smaller / weaker models basically invariably struggle.. when given a bunch of somewhat unrelated questions / tasks dumped on it.. whereas stronger models generally seem more capable at working through it systematically and accurately

[ocean vortex]

honestly this is kinda decent for lmarena specifically to test many things fast and not flood the context that you cannot clear there. How did 'dreamtides' do on this?

[alpine coral]

aha it was borne out of lmarena

i haven't had dreamtides

[keen beacon]

are there models who have gog this?

[sonic tendon]

wait, how was this found lol

[calm sequoia]

It would be such a self-sabotage if Chinese models would use SciHub database and westerners would be banned from that

[eager mica]

Llama 4 got released in spite of the Kadrey v. Meta Platforms copyright lawsuit (from which it was found that Meta used pirated books beyond the Books3 dataset in the pretraining data—not that other AI labs aren't doing it too) which is still continuing. Makes me wonder if some of the data was taken out of the publicly released models.

[keen beacon]

interesting

got* lmao sorry

[balmy mist]

Wow bro this is so clean, I’m interested to know your prompt thought process for so the stages

How is dreamtides?

And all these names man lol

[sonic tendon]

annas-archive + scidb, more like

[balmy mist]

the quasar model is free from openrouter api right?

gotta make sure before i start stress testign with it

google is my fav company now, like they got a free SOTA model and they give you $300 free credits once you get rate limited on free plan

like they really gained a fan of me

fr man and you literally have studio

where you can just use it unlimited pretty much

crazy

but Anthropic to much money

yeah bro

[cedar tide]

What happened to the Luca model?

[balmy mist]

they were the first to do the branching

like man

have yall tried having like 6 plus tabs of studio doing different generations at the same time?

the fact that we can do that is nuts lol for free lmaoo

[balmy mist]

which one was that

[cedar tide]

Write Luca in the search bar

[balmy mist]

of googel?

[cedar tide]

Here

[balmy mist]

lol

was it a good model? we getting a lot of chinese models

i heard someone say last week that nw dropping today

[cedar tide]

What is nw ?

[cedar tide]

Who say that ?

[balmy mist]

someone on twitter

let me find post

[cedar tide]

The google next cloud 25
start tomorrow

[balmy mist]

oh its tmw

damn

its okay im cool with gemini 2.5 fr

[cedar tide]

Its from the 9 to the 11

[balmy mist]

i happy for new models but workign with 2.5 has been so good

[keen beacon]

stargazer/lunarcall/dreamtide one of them will probably drop tmrw/this week ig (2.5 thinking models)

[balmy mist]

but i still want nw lmaooo

[balmy mist]

which one of those are best

i didnt try dream

[keen beacon]

stargazer i think

[balmy mist]

eww

im not a fan of star

[keen beacon]

there might be the same model different revision in those anon names too

[balmy mist]

lmaooo

fr

[keen beacon]

yeah

[plain zinc]

https://x.com/AdvaitOnline/status/1909577360875143633

Advait Bopardikar (@AdvaitOnline) on X

♊🤖🎰✨🛳️⚡📈

X · Apr 8, 2025, 12:01 PM

Tomorrow

[balmy mist]

yo i love that they be responding to our requests on twitter lol

im tryna think of what is missing from studio

the only thing i would say is making the UI look more cute but that stuff doesnt matter as much

[keen beacon]

its obviously supposed to be openai

they say openbrain model spec and then link to openai model spec

[cedar tide]

https://x.com/reach_vb/status/1909584596401815691?t=2s9zE6jxcor6uvc2r2835w&s=19

Vaibhav (VB) Srivastav (@reach_vb) on X

NEW: Llama 3.1 Nemotron Ultra 253B - beats Llama 4 Behemouth, Maverick & competitive with DeepSeek R1 - Commercially permissive! 🔥🔥🔥

Open weights on the hub!

X · Apr 8, 2025, 12:30 PM

[keen beacon]

its the authors' intention it might not play out that way tho

[calm sequoia]

The upcoming o3 is not hyped enough

And I am startin to understand the hype on the deepseek

https://x.com/HamedMahdavi93/status/1908198747684680181

Hamed Mahdavi (@HamedMahdavi93) on X

🧵 1/ How well do LLMs actually do on Olympiad-level math?

We evaluated frontier models on 455 problems from the IMO Shortlist.

Unlike most benchmarks, we emphasize proof validity, not just final answer correctness.
Here’s what we found 👇

X · Apr 4, 2025, 4:43 PM

[balmy mist]

im actually a lil exicted to try r2

i hope they launch this week

want to try it with my new app lol

[calm sequoia]

Is there info for R2 to be launched this week? I though Alibaba model will drop and not R2?

[keen beacon]

So is it true nightwhisper might be the top model specifically in terms of coding or in general

[calm sequoia]

In "general" nothing beats 2.5 Pro as far as I know

[keen beacon]

What about coding?

Does NW out do it there?

[balmy mist]

NW is the king until proven otherwise

[keen beacon]

Yes

[balmy mist]

gemini 2.5 is the best overal model tho, we need to get our hands on NW outside of webdev

[keen beacon]

So currently its unavailble right? @balmy mist

[keen beacon]

Yes as an agent ide, cant wait

[balmy mist]

it was performing at the same level as gemini 2.5 pro for general stuff for me, but we need it outside of webdev for better tests

yea we cant use it

[keen beacon]

must be cap

[balmy mist]

why?

i mean you are right

there is no reason for them to release it

i would wait if i was them

see what other slaunch

and then launch it after

they can make it better and wait

[keen beacon]

NW is a game changer in coding as its demonstrated on the arena. Even in its beta.

[balmy mist]

i would tho keep puttign it on webdev and lmarena under diff names every once in a while lol

its funny it was only there for like 2-3 days

manyeb only 2 actually

and they took it down so fast

[keen beacon]

So when will it be an estimated preview/release as an agent that can be used in an IDE like cursor?

[balmy mist]

actually mayb 1-2

3.7

easily

but to me im not paying for that

[keen beacon]

Also, web arena says claude 3.7 sonnet beats gemeni in coding, is that true?

[balmy mist]

you cant justify me doing that wen u have gemini 2.5 for free

[balmy mist]

preference thing imo

the difference is small tbh

so its a lot of preferences

pokemon games lol

[keen beacon]

lol

[balmy mist]

website for my projects

[keen beacon]

nothing

What model do you guys recommend thats most effiecent in Java coding?

[balmy mist]

these one is an arena for llm agents to play games against each other
one is my website for my small scale apps like pokemon games
one is a big game
one is a app to run a bunch of iterations on a model

one is a matrix sim
one is a story teller ai

most not fiinsihed tho lmaoo

only like 4 is finished

[keen beacon]

You said they create games, or play them?

[sonic tendon]

i feel like o3 might start trialing on lmarena soon

[balmy mist]

play games, rn i only have maris kart sim

[keen beacon]

Does it use an algorithm to learn to play?

[balmy mist]

i wish bro

[sonic tendon]

seems like the only model that has a solid chance at beating gp2.5 (and maybe r2)

[balmy mist]

you got a screen recording of it?

im trying to make mine better

[sonic tendon]

how are you guys so well-connected 😭

[balmy mist]

wait wait

can you run some of our prompts?

[sonic tendon]

oh yeah i was gonna ask

[balmy mist]

do the pokemon test please

[sonic tendon]

and the jar test

[balmy mist]

its funny we still dont have grok api

once you start coding with ai like using the api you can never go back

[sonic tendon]

reverse-engineer lmarena api 💯

[balmy mist]

its so much fun

[sonic tendon]

wait, fr?

where's that

[keen beacon]

there isnt afaik

[keen fulcrum]

What do you think? https://www.plaud.ai

[fierce sinew]

Is https://github.com/lm-sys/FastChat still maintained? Would any of the current maintainers be open to bringing on a new maintainer (our team would be happy to volunteer). Who would be the right person to talk about this?

[keen beacon]

@hollow ivy is this true?

[sonic tendon]

why not just use your phone's recording app and then transcribe it

i think both stock android and ios do that automatically

[keen fulcrum]

Thats not a great microphone

[sonic tendon]

plus, wearing a wiretap 24/7 is sorta creepy lol (sorta /j)

[keen fulcrum]

And it uses AI while recording

[sonic tendon]

??? what

[balmy mist]

whats teh fastest model that is free? like is there a fastest model then gemini 2.5?

[keen beacon]

for anyone who sent me prompt requests for the private model just dm me your prompts pls

[sonic tendon]

cool site btw

[keen beacon]

worked on it for most of yesterday, thanks!

[balmy mist]

you tested simple bench already right?

from here?

https://github.com/simple-bench/SimpleBench/blob/main/simple_bench_public_set.csv

[sonic tendon]

the poetry-type stuff is cool and kind of a mood

not sure what to call it

[keen beacon]

i used the try it yourself set on their site

it got 6/10, which is the best score of any model on that set iirc

[keen beacon]

haven't tried the 20 set

don't currently have time to do it but will see

[balmy mist]

nice thanks bro, i tested all models so far with pokemon 0-shot

and nw had the best followed by gemini 2.5 and sonnet

[keen beacon]

in the testing i've been doing with this

it is definitely better than o1 and o3 mini at web tasks, but gemini 2.5 pro and claude 3.7 sonnet are still better

just seems to be something oai

don't have the best data for

[balmy mist]

damn

but thats still progress

but to score 6/10 is impressive

you are not using the pro version of that model right?

cause o1 pro is so much better than o1

[keen beacon]

nope

my current hypothesis is this is o3-medium

[hardy pecan]

how did you get access to the model?

[keen beacon]

can't say

[hardy pecan]

ok..

[visual turret]

What is dreamtides

[hardy pecan]

Google model, its meh

[keen beacon]

unreleased 2.5 thinking model

[kind cloud]

maybe flash-lite

[visual turret]

Maybe 2.5 pro non reasoning

[keen beacon]

it isnt it thinks

[hardy pecan]

yeah, shes a thinker

[kind cloud]

At least, dreamtides is weaker than flash-thinking-01-21 in math, as far as I know.

[visual turret]

How do you know

[keen beacon]

against a non thinking model in a battle it waits for the thinking model to think first

u are not waiting a minute for the first token on a non thinking model

[keen beacon]

u have to test it with something that requires a lot of thinking thats when its most obvious

[visual turret]

Tbh anthropic research shows it makes the model lie more

[keen beacon]

because the thoughts can be very short/fast depending on the model

ok but thats unrelated lol

[visual turret]

It isn't

Anthropic did research on reasoning models

And they found it made the model lie more

[keen beacon]

we are talking about how to know whether a model in the arena is thinking or not

[visual turret]

We aren't

Your losing an argument and your trying to make it be about something you know

[keen beacon]

????

when was i talking about anything about reasoning models and lying lol 🤣 🤣 🤣

i also just measured math 500 for quasar so benchmarks for it:

gpqa diamond: 67.42%
math 500: 90%

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

yeah so it looks like a 4o update

another one 💀

[ocean vortex]

it's kinda big this time as it's finally gonna be dated API model release with metrics. And looks like 1M context too

so it's not gonna disappear after they update it lol

[keen beacon]

try it on aistudio with web search enabled

[subtle flower]

I am trying to compare Gemini 2.5 Pro and gpt-4o in the side-by-side and why does Gemini stop generation as soon as gpt-4o generation is done? It seems to stop in the middle of it's answer

[drifting thorn]

god I've just receive a code from Manus!!!!!

Just 18 hours before when I checked my email

[balmy mist]

i used it once and never went back

its good for like basic stuff

[barren prairie]

On arena battle this works ...only on side by side or direct chat Gemini anwser always cutted

[balmy mist]

or research

but the hype for it died down

they should have gave more ppl codes early on

now we all kinda moved on and their are so many other versions of it now

but im still happy i got a code lol

but im not tryna pay for that lol

[drifting thorn]

genspark heard to perform better than Manus, but Genspark has a low token limit

[balmy mist]

yeah true

[drifting thorn]

It basically forces you to pay

for the token

[balmy mist]

yupp same with manus after like two tasks

i used 600 credits of the 1000 free ones on one task lol

[drifting thorn]

And I think my task has a very long context but the further action is rather simple

most rag-based llm just failed because they run out of context limit

for example Flowith Gemini 2.5 Pro with knowledge base garden

it failed, saying "context too long"

[balmy mist]

anybody have any prompts for me test on my app that recursively iterates on it? i am using a webdev system prompt so prompts like that will be great

also tell me how many refinements you want, this will be using quasar

[balmy mist]

what app you used when you used gemini?

[drifting thorn]

flowith, ive told ya

also cherry studio doesn't work

cline performed the worst

there's no option for knowledge base in api studio

basically i only think multi-agent will work for my task rn

[balmy mist]

oh yeah

i think you gotta prompt it right

and use system prompts for agents

[keen beacon]

gpqa diamond: 71.4%
math 500: 87.1%

[drifting thorn]

Like, I've been writing a fanfic for 80 chapters with Gemini 2.5 Pro, when 2.5 Pro is able to summarize characters in a novel well. Then, it goes 80 chapters, and 2.5 Pro is not able to hold up the details anymore

since summarization means losing details

[balmy mist]

why does everyone like qwen so much? is it really a good model?

[drifting thorn]

an interesting model for QwQ 32B

it 'thinks' a lot

[keen beacon]

qwen are a team theyre about to release their own llama 4 analogue qwen 3 soon

qwq 32b matches r1/etc (much larger) in rote tasks in my experience and its based on a base model released in september2024

[balmy mist]

hmm you got to playe around with the how you are summarizing, thats the key, its not about all the details but the right ones that enough for the llm to make new chapts and you have to guide it with what you want the next chaps to be around and provide the missing details

but hey i never did anything like that lmaoo

but there is an app for this

@drifting thorn try this video, might be usful for you: https://www.youtube.com/watch?v=MBcA4iaQs_M&ab_channel=MattVidProAI

yeah i cant at him

[balmy mist]

is qwen usally fast? how much are their in and outs?

also the context has to be 1 mill right?

the only thing that matters at this point is 1 mill context, cheap ins and outs, and good output amount, also speed, i think i have a solution for IQ with my setup

but i need it to be fast inference

like faster than quasar

[keen beacon]

i mean they have an api but most people dont use it. u can run their models locally unlike llama 4 which is too large. most people use other providers like together/etc so it really depends on how other providers price them

[balmy mist]

but cant be dumb lik llama4 tho lmaoo

ill prob use openrouter

i need to buy a setup just for ai lol

[keen beacon]

[drifting thorn]

This doesn't seem to be useful

Since I have a looooooooooong chunk of different fictions

[keen beacon]

most models cant do 1m context even if they support it. this also applies to gemini on a lot of tasks. for doc summarization, etc., tasks that are in distribution they can do it though.
128k context is enough for me personally, if its done well and works on a lot more tasks

i don't know what proof you want.. i'm not allowed to share screenshots or directly ask the model, but i can take prompts

[drifting thorn]

and the interface is just not my thingy

[keen beacon]

"i'm not allowed [... to] directly ask the model"

i will get flagged

the point is that it is a private model

but from my testing

it thinks, is similar in style to o1, and performs better in my tests

[balmy mist]

@keen beacon you can give us outputs? like for the pokemon thing?

if not can do a screenrecording of the game on liveweave

[balmy mist]

damn sorry, i thought that would help, maybe you gotta create your own app

thats what i would do

[keen beacon]

running the prompt now

o3 medium output

cc @balmy mist

[balmy mist]

thank you so much bro!!!

[keen beacon]

(replace the "???" with "meta" btw, the frontend replaces all ai lab names with question marks to prevent identities being learnt)

only o3 medium

i think o3 high/pro will be quite a lot better

as is the case with o1 & o3 mini

[balmy mist]

its pretty fast tho, the output is okay, yeah o3 pro will be goat

[keen beacon]

no

i don't work there

i just help them out sometimes 😉

oh yeah

another thing i've observed

it is noticeably better at language-related problems, something o1 was already strong in

it feels more clever, even for someone who isn't well versed in the thing they're asking about

it also thinks longer in general but not excessively so

2.5 pro still better for code though 😔 will have to see what o3 pro can do

yeah there's a balance

[balmy mist]

has anyone here used groq api? im trying to play around with it, also has anyone used their app gen thing?

[keen ferry]

I had manus overloaded and i got my 900 credits back lol

[drifting thorn]

Is that 1000 credits one-off or replenished daily?

[balmy mist]

one off

they said that they will give extra every now and then

but you gotta sub

to get more tbh

yo groq truly is fast af

after 15 refinements:
this from o3 medium

to this being refined by quasar:

if I refined with gemini it would have been a lot better

[keen beacon]

how did you get the model to refine

a generic prompt or did you point out specifics?

[balmy mist]

specific prompt, check that system prompt i gave you, pretty much that

but i have a bunch of ones like that

refactor

and when you play with the temp its get even wilder

[balmy mist]

you can also point out specifics as well when you give the initial prompt, its all depens on your system prompt, if you want i can open source it so you can try it out?

i wanna make a dark mode next but was having trouble with it

if yall got test prompts for me let me know, i am curious to see what the llms cooks up

[drifting thorn]

Is Grok 3 sensitive to “bad prompts”?

[balmy mist]

honesly i think the companies are doing this behind the scenes lol

[keen beacon]

looks cool

[balmy mist]

thanks bro, i dont think i should have asked that prompt lol

i thought it would be quick

[keen beacon]

one thing i like doing just for fun is asking new models what their spotify playlist would look like

o3's:

Here’s how my personal “AI‑Curated Grooves” playlist would shape up on Spotify. I aimed for a start‑to‑finish arc that moves from bright, feel‑good energy to late‑night wind‑down—all while hopping genres to keep things interesting.

Playlist title:
AI‑Curated Grooves: Sunrise to After‑Hours

Short description (what you’d read on Spotify):
An eclectic ride from morning pep to midnight calm—indie pop, hip‑hop, alt‑R&B, classic cuts, and a few global gems.

Cover art concept:
A softly saturated city skyline at dawn, with a subtle waveform running across the sky.

Track list (in play order):

1. Levitating – Dua Lipa (3:24)
2. Dreams – Fleetwood Mac (4:17)
3. Heat Waves – Glass Animals (3:59)
4. Electric Feel – MGMT (3:49)
5. Sundress – A$AP Rocky (3:22)
6. Lost in Yesterday – Tame Impala (4:09)
7. Blinding Lights – The Weeknd (3:22)
8. Coffee – beabadoobee (2:14)
9. Take a Chance – The Magic Gang (3:20)
10. Kyoto – Phoebe Bridgers (3:05)
11. Good Life – Kanye West feat. T‑Pain (3:27)
12. 2009 – Mac Miller (3:56)
13. Nights – Frank Ocean (5:07)
14. Borderline – Tame Impala (4:34)
15. Borderline (Anitta Remix) – Anitta & Miley Cyrus (2:51)
16. Teardrop – Massive Attack (5:30)
17. Pink + White – Frank Ocean (3:04)
18. Cherry Blossom – RINI (3:33)
19. Slow Dancing in a Burning Room – John Mayer (4:02)
20. Moon – BTS (3:29)
21. Plastic Plants – Mahalia (4:16)
22. Peach – Kevin Abstract (3:30)
23. Fade Into You – Mazzy Star (4:55)
24. Motion Picture Soundtrack (strings version) – Radiohead (3:20)
25. I Know the End – Phoebe Bridgers (5:47)

A few notes on the flow:

• Tracks 1–7: High energy, bright synths, and upbeat rhythms—great for morning commutes.
• Tracks 8–13: Indie and alt vibes to keep focus during work hours.
• Tracks 14–18: A dip into hazy psych‑pop and downtempo electronic for early evening relaxation.
• Tracks 19–25: Guitar‑driven moodiness and cinematic closers for late‑night introspection.

Total runtime: ~1 hour 35 minutes.

Feel free to tailor the order or swap songs based on your mood—Spotify’s radio and “Enhance” features can also help expand the vibe further. Enjoy the listen!

there are some misses on here (very overplayed pop songs) but also some bangers

tame impala, frank ocean, radiohead

[balmy mist]

that is so interesting lol

[keen beacon]

i think chatgpt 4o 20250326 had the best one for my taste

1. Radiohead – Weird Fishes/Arpeggi
   Hauntingly beautiful; deep dive into sound.

2. Phoebe Bridgers – Motion Sickness
   Indie melancholy with sharp lyrical wit.

3. Bonobo – Kiara
   Organic meets electronic—perfect thinking music.

4. Tame Impala – Let It Happen
   A psychedelic groove that just keeps going.

5. Frank Ocean – Pink + White
   Soft, emotional, and timelessly smooth.

6. Arctic Monkeys – Do I Wanna Know?
   For when overthinking turns sonic.

7. The Cinematic Orchestra – To Build a Home
   That one song that gives you goosebumps, every time.

8. ODESZA – Across The Room (feat. Leon Bridges)
   Uplifting yet grounded in soul.

9. Billie Eilish – when the party’s over
   Minimal and powerful.

10. Tycho – Awake
    For daydreaming while watching the world go by.

11. Sufjan Stevens – Mystery of Love
    Delicate and bittersweet storytelling.

12. Massive Attack – Teardrop
    Atmospheric with a pulse.

13. Fleetwood Mac – Dreams
    Timeless groove and emotional resonance.

14. James Blake – Retrograde
    Layered vocals and soul-tugging soundscapes.

15. Lana Del Rey – Mariners Apartment Complex
    Romantic, melancholic, and bold.

weird fishes, let it happen, do i wanna know are all great

oh yeah i haven't tried it with quasar alpha yet

similar to 4o latest, pretty good

fire stretch

[balmy mist]

wow frank ovean and billie ellish

bangers

i like 4o too

ask gemini and sonnet

[drifting thorn]

So sad I’m a K-pop listener

[balmy mist]

dont be sad, music is diverse

[drifting thorn]

None of the above is sung by a girl group

[balmy mist]

thats what makes it special

[drifting thorn]

You’re right

[keen beacon]

#general message o3 included BTS

[balmy mist]

what about refining music hmm

need a new system prompt

@hollow ivy why you didnt tell llm to give you straight up midi?

[keen beacon]

c3.7s thinking playlist

beach house, radiohead, bon iver, aphex twin, kate bush, radiohead

this one may be my fav

there will definitely be a release tomorrow

and i do think chances are it will be whatever nightwhisperer was

either gemini-coder or a non-preview release of 2.5 pro

[balmy mist]

omgg that would make my whole week

might have to take off work the rest of week

[ancient reef]

updated

[balmy mist]

noooo

please say it wil be nightwhisper

thats my wife

she ran away from alter tho

been trying to find her since

what if there never was a nightwhisper

it was just apart of our imagination

[fleet lintel]

they wont launch a new model. it will be part of new version of 2.5 (pro or flash). I doubt they will have coding specific model

[balmy mist]

https://x.com/minchoi/status/1909623992454938624

Min Choi (@minchoi) on X

This is wild.

DeepSite is like Cursor AI but in your browser, powered by DeepSeek V3 0324.

Now you can vibe code AI apps, games & landing pages.

100% free

5 wild examples (and how to try it yourself):

1. Cyber Flappy Bird clone

X · Apr 8, 2025, 3:06 PM

[keen beacon]

they have historically been fairly significant jumps iirc

particularly for 2.0 flash thinking

[balmy mist]

this dope af

similar to bolt

but a deepseek version lol

[keen beacon]

i agree that it's not great, but that wasn't my point

my point is that preview vs full releases from google have still been significant jumps and i wouldn't be surprised if the same goes for 2.5 pro

?

[fleet lintel]

it's bad now. 4 months back, it was decent-ish 🙂

[keen beacon]

that happens with a lot of models.. if you regen enough it'll get it wrong eventually, but that doesn't mean it always gets it wrong

[lime coral]

I mean it always depends on what you are trying to do

Definitely not here for coding

[fleet lintel]

how do I understand all the openai model terminology. there is o3, o1, mini versions, gpt 4, 4.5 etc. What is what ?

[lime coral]

But people like it for retrieval in audio/video and sometimes writtingn

I am not a flash/mini person. I only love and use big gun. This is why i am sad no ultra since its release

[fleet lintel]

that's helpful.. And increasing O means better models? Like o3>o1 and so on?

[keen beacon]

i miss ultra too 🤝

[fleet lintel]

thanks. And which models are present to whom? how do I know? like free chatgpt is o3-mini? and 20$ is what and 200$ is what?

[keen beacon]

it was fantastic at creative writing

dense 1T+ param models will never be beat language wise

but for everything else they're impractical

[fleet lintel]

may be not enough money and demand for ultra models?

[brittle tiger]

Seems like a lot of GDM folks hype posting (rare for them) about tomorrow which I don't think they'd be doing if it was going to be underwhelming

[fleet lintel]

links?

[balmy mist]

nahh this deepseek stuff is fire, i didnt even know you can use hugging fac like that

[fleet lintel]

and tomorrow is cloud event but businesses. i dont think it will be much interesting for users.

[fleet lintel]

I just tried couple of small games in deepsite.. it's kinda amazing!

[balmy mist]

yall gotta try out that deepsite, its like nightwhisper

fr

maybe nightwhisper is just bunch of agents?

im getting the same results i did with nw

but it takes a longer time, but this app can be deployed on a site

easy to share

and store

wild

[fleet lintel]

i think nightwhisper is still slightly better. but for opensource model, deepseek is killing things

[balmy mist]

idk bro

for one shot?

[drifting thorn]

I hate mini models by the fact that mini models has a worse chain-of-thought than big models

[balmy mist]

night whisper is the best stand alone, but this platform is optimizing the model using a bunch of tools

if nightwhipser had that it would be cracked cracked

im just saying its producing the same level that nightwhisper did with how they implemented it

[leaden palm]

their main page isn't

[balmy mist]

imma try and mimic this

imma show you my one shot pokemon game with this, it shows its not about the model as much once you get to a certain level of inteligence

[drifting thorn]

So sad that Manus has a subpar base model, making the writing looks back

[balmy mist]

deepseek r1 is the miminum or maybe v3.1 not sure what they are using

[leaden palm]

still literal misinformation

[balmy mist]

but prob sonnet 3.5 level models and above are all you need going forward

just have to prompt it right for system prompts and give it tools

thats why you see manus

[drifting thorn]

Idk if I’ll wait for 2.5 Pro to open up the 2 million token context window

[balmy mist]

and all these other stuff

nahh deepsite is cracked, i found my new baby

hugging face pro is $9 wow

cracked

yo

im a feen now

[leaden palm]

ahh this was from https://x.com/ArtificialAnlys/status/1909624239747182989

Artificial Analysis (@ArtificialAnlys) on X

Llama 4 Intelligence Index Update: We have now replicated Meta’s claimed values for MMLU Pro and GPQA Diamond, pushing our Intelligence Index scores for both Scout and Maverick higher

Key update details:
➤ We noted in our first post 48 hours ago that we noticed discrepancies

X · Apr 8, 2025, 3:07 PM

[fleet lintel]

I dont understand why Meta is not able to compete with deepseek? they have probably lot more resources in both Engineering and Machines.

[fleet lintel]

i am out of free limits

[eager mica]

"Safety", "copyright", too much red tape, too many cooks on the project.

DeepSeek R1 was considered one of the "most unsafe" SOTA models available when it came out.

DeepSeek didn't care about that, and nobody could do anything about it (well, somebody tried).

[fleet lintel]

safety is overrated right now

[balmy mist]

lmaoo already?

[fleet lintel]

things are improving too fast

[balmy mist]

im about to hit mine as well

i think what you should do is start there and then offload it to gemini

it made this:
https://huggingface.co/spaces/IjedMeer/test_run

[drifting thorn]

Open-weight models can be jailbroken

[balmy mist]

but it is not playable, got to fix it with gemini

[fleet lintel]

share the prompt

[balmy mist]

i gave it my code from an existing game i made lol

so cheated a lil

but it was a game made by gemini

like with 3 iterations

meaning i asked make pokemon game

then grabbed the output and used it as input and said make it better

did that 2 more times

and thats what I put into deepsite

now i have gemini fixing it, i will update the code as soon as gemini is done and it should be playable again

oh the prompt i used was make it 100x better for deepsite lmao

next time imma say 1000x

its fun to see how they interpret that

but you can see the site when you click that link right?

[fleet lintel]

yes.. it just say initializing

[balmy mist]

omgg

cooked

updating now

just got to fix one thing that is not letting me deploy

vibe coding 101 lmaoo

[fleet lintel]

i think 3.7/3.5 are good for small projects. comparable to gemini. But 2.5 is better for bigger more complex projects. Given the cost, i am switching to gemini over Claude

[teal mantle]

put some funds on openrouter

should I regret it?

[keen beacon]

If u put in 10 USD u can use quasar for free without rpd

[teal mantle]

I did

what is rpd?

but quasar gave 502

[keen beacon]

Requests per day

[keen beacon]

Maybe a region thing?

Free and basically unlimited gpt 4o api for the time being lol

[teal mantle]

ah

protonvpn worked in allowing quasar access

this sucks

[keen beacon]

Ya they added the restriction yesterday

Well u can do a lot more on quasar for now

[teal mantle]

I kinda envy you having access to vpn vanilla, or living in somewhere that do

even using quasar I just realized I need VPN

for Christ's sake

[leaden palm]

maybe running benchmarks or generating synthetic data if you like that

[teal mantle]

I mean API vanilla

without using any third parties

Quasar is definitely OpenAI

lived in banned areas
the pariah state of artificial intelligence

[keen beacon]

Quasar is just an updated gpt 4o

[teal mantle]

Based on stylistics you're probably right

[leaden palm]

what is "api vanilla"

[teal mantle]

But for openrouter how should I spend the credits on?

[leaden palm]

are you trying to use "vanilla" to mean the sense of cleanly/simply?

[teal mantle]

using just the API without any third party

for example, not using openrouter

[leaden palm]

last time i checked openrouter doesn't get around geoblocks

[teal mantle]

I thought it does
let me check again

but man, OpenAI sucks for geoblock

[keen beacon]

Maybe just spam quasar for now lol

U can decide later

[teal mantle]

I live in a defacto geoblocked area though

[leaden palm]

average reaction to geoblocking

[keen beacon]

1 per sec no concurrent limits I think

[teal mantle]

an average reaction towards the notion of being blocked because you somehow want to keep in touch with tech in general

[keen beacon]

Actually if u have 10$ now which they require for no rpd u can do 10 req per sec hmm

[teal mantle]

@leaden palm why wouldn't geoblock suck? too bad this is a Christian server, no swearing or else you know

[keen beacon]

It's free tho lol

[tawdry meteor]

Which one was NW? I missed it when available, so hard to keep track 😅

[balmy mist]

wait i been using quasar non stop, wym i have to put 10 in ? @keen beacon

[leaden palm]

no

it works fine

you just might get higher limits if you have $10 in active balance

[balmy mist]

hmm okay

ill test that

[teal mantle]

just one question
do most of you mind if your data's used for training esp for Quasar

[balmy mist]

lmaoo

im not puttong personal dara in

im making pokemon games

[keen beacon]

There's 1000 rpd limit if you don't have 10 dollars in

[balmy mist]

ahhh

[keen beacon]

No if I'm doing 100m tokens an hour

[balmy mist]

bet i will put 10 in and do 2000 requests

[teal mantle]

for real though

based

[balmy mist]

i need to stress test this app

just did refinement of pokemon game from deepsite with quasar 20 requests in 1010 seconds
w/ each request containing around 30k tokens of code input and output give or take

yall positive its unlimited with 10 in your open router?

also someone give me a easy coding prompt to run 100 times

[torn mantle]

nice

[keen beacon]

Alex said it's subject to supply here #1357398117749756017 message (presumably unlimited like it was before)

Under 10 and u only get 1000 rpd

[balmy mist]

he works there?

[keen beacon]

He's the owner of openrouter

[teal mantle]

do you need to have the 10 dollars unused or what

[balmy mist]

oh i see

[keen beacon]

Ya I think

[balmy mist]

thnx

[keen beacon]

You get higher rps with more credits anyway

[balmy mist]

true

so best model on openrouter is quasar now?

[keen beacon]

Free frontier model that you can spam api requests with

[balmy mist]

i like that

this is perfect for me

gonna go ape on it

i just wish we had higher outputs

[leaden palm]

well it's 4o with long context and free

not 4.5 level model

[balmy mist]

thats the only thing holding it back

[leaden palm]

not a reasoning model

[balmy mist]

fr you cant beat free

nd its a better 4o

and fast as a mother

nahh bro you wild

grok dont get respect until they release api

like y have they not done that?

makes no sense

very sus

openai is expensive

but they still released it

it bad for market imo

but give us the option

why limit us to their platform

[teal mantle]

reasoning market isn't as hot as the o1-o3 days

[balmy mist]

didnt he found openai on opensource goals?

[teal mantle]

nonreasonings become popular again at 4.5

[leaden palm]

there's a reason 2.5 got hype and 2.0 didn't

[teal mantle]

hence the suits

[teal mantle]

yeah, except gemini

[leaden palm]

?

everyone is making reasoning models

deepseek

nvidia

anthropic

soon meta

reason is only getting better

reasoning is the new scaling paradigm

[teal mantle]

rather i mean the temporary change in you know flavor of the week

[leaden palm]

we hit a wall, and we will climb it with reasoning
*or at least RL

[teal mantle]

advancement is a good paradigm, just flavor's alignment is the question

my point: reasoners are the future, but not every time the future is the flavor of the week

[balmy mist]

okay im trusting you @keen beacon gonna do a 100 run refinement

[teal mantle]

people thought we did during gpt 4.5 early weeks

[leaden palm]

well we have their word, and a few other ones like livebench iirc~

nvm livebench doesn't have them

but there are some other independent benchmarks that include grok 3 iirc

would be great if i could find them...

if it was expensive i would expect them to have tighter limits

[keen beacon]

It is pretty weird

[leaden palm]

(unless they just have so many gpus they might as well use them)

[ocean vortex]

disagree. There are some things gpt4.5 does better than updated gpt4o but there are MORE things where the opposite is true. It's kinda like comparing 3.5 sonnet with 3.0 opus

[leaden palm]

do you think it is the best model on OR then?

[ocean vortex]

it's not obviously, but it's not reasoning model. Among non-reasoning models it is now one of the best I believe yeah

[leaden palm]

hm

[ocean vortex]

depends what you want doing with it too. I wouldn't use it for web or design development but for everything else where you don't want reasoning it's a great model now tbh

you could to some relatively small extent but it was not trained for it. So like 30k responses are very not possible lol

the most I got from it I think was like 4k tokens

we should test it on simpleqa. Assuming openai continue refusing to release it officially

[lime coral]

Tomorrow the party begins. At least two players. As long as Google moves OAI follows

[balmy mist]

this is the 39th refinement of the 100 refinements:
https://huggingface.co/spaces/IjedMeer/test-app

not bad so far

yall should see the first one

cant wait to see 100th lmaoo

just hope its in a working state becaue the 14th was not

if 100 dont work ill keep going down until i find one that does

this is like a cheat code for getting an idea to a fully fleshed out version, just gotta wait lmaoo

with one prompt you can get a solid game:
make the best snake game that has a bunch features and sounds and great visuals

the sounds dont work, but hopefully the 100th version of it does

if yall have any other test prompts for me to try let me know

[cloud meadow]

It's got an embed now?

https://alpha.lmarena.ai

[golden ocean]

no, it's an optical illusion

[cloud meadow]

https://tenor.com/view/oye-cat-enough-stop-petting-gif-8639785326678511628

[keen beacon]

KITTTTTTTTY

[cloud meadow]

😼

[keen beacon]

[brittle tiger]

I'll post some as I see them. This one isn't hype posting and I'm for sure overanalyzing but I follow AI researcher Twitter very closely and vibes from GDM side seem at all time highs

https://x.com/epiqueras1/status/1909666650422743145

Enrique Piqueras (@epiqueras1) on X

Feeling grateful.

https://t.co/g0DCQxJbjY

X · Apr 8, 2025, 5:56 PM

[primal orbit]

guys, based on those 2000 prompts they published on maverick, did you manage to establish which model it was when it was anonymous? I see someone wrote it was spider.

i can't find my prompts which i send to 24 karat, so I assume it's not it.

did anyone find his own prompts?

https://huggingface.co/spaces/lmarena-ai/Llama-4-Maverick-03-26-Experimental_battles

[ocean vortex]

this is only using models that are available in direct chat I think

[spare mango]

bruh gemini is free and better than paid chatgpt and no one has a clue about it.

except, ofcourse, those that keep uptodate with LMArena.

[ocean vortex]

yeah that's batsht hilarious that it has a decent chance for a better response than o1-pro can give you for $10+ per request. Completely for free like it;s just a thing to do LOL

[spare mango]

yeah that should be criminal, I have no idea why this isn't common knowledge yet.

How is this fact being kept hidden from the masses even though it's not private information?

[ocean vortex]

I think much more people are using it now than in the past. But yeah some of it is seemingly deliberate marketing by google

they are not pushing AI from their main website

you just have to know or hear about it elsewhere. Despite google.com being like the most traffic generating website 👀

[spare mango]

That is so weird isn't it?

[ocean vortex]

google ads. They need to make up their mind finally and go full in

[spare mango]

Wonder if there's some kind of partnership of Alphabet with OpenAI to keep this information relatively unknown to the average person.

[ocean vortex]

nah absolutely not they are competitors lol

but google is making money from ads

that only really work with outdated seach and no real AI/gemini

[spare mango]

Yeah so why aren't they going all in? They can make big bank with AI, given they have the highest-of-the-line product available.

And consistently have the best product available in the market for a while now.

First it was 2.0 thought experimental

[ocean vortex]

they could and they should. At this point it seems that's kinda inevitable either way. That classic search is gonna become legacy and not very relevant sooner or later

[spare mango]

Now it's 2.5 Pro

[spare mango]

Yeah so why haven't they?

[ocean vortex]

investors and shareholders is my guess

[spare mango]

Google being a tech monolith, surely would not sit on the fence for massive financial decisions such as this.

[ocean vortex]

all that bureaucracy of a big corp and having teams that sometimes work against one another

[spare mango]

Hmm could be...

[lime coral]

https://x.com/amrmalameen/status/1909688535407640623?s=46

Amr Khalifa (@AmrMAlameen) on X

@lepikhin @scaling01 🤫

X · Apr 8, 2025, 7:23 PM

Both are from DeepMind

What if ultra is nighwhisper and the other dude wasn’t trolling it would be so funny haha

[golden ocean]

true

[torn mantle]

wtf

no way

but why would they test it on webdev and not lmarena?

[lime coral]

lol I actually don’t think nightwhisper is ultra it was a joke, but it’s pretty obvious now that ultra is coming don’t know when though

[balmy mist]

Agreed

[lime coral]

My guess is at the I/o or June and December/November for Gemini 3 with GPT5 in August

[torn mantle]

yea probably

im so for a gemini coding model

it will be cost efficient

and more affordable

instead of gemini ultra

[balmy mist]

Me too

[torn mantle]

tomorrow gonna be lit

[lime coral]

Ready to use ultra for free in ai studio. If they allow it, I might feel like I am stealing something

[torn mantle]

you dont strike me as the type to pay for anything tbh

[barren prairie]

I don t think that google will stay closed hands all this periode 😆

Maybe we will have Gemini 4

[balmy mist]

It seems like Google has found a way to quickly train models and get them released fast

Like they had a slow start but with that foundation they built it seems efficient as hell

[torn mantle]

infinite TPUs

they are improving on both SW&HW

[oblique flint]

cant tell if this is bs or not. Seems a bit odd naming for google

[vivid oyster]

It means they're gonna make a model for coding but its not gonna be called gemini coder 1

[golden ocean]

fr

phil getting his leaks from mcdonalds

(idk what i'm talking about)

[cloud meadow]

You'll say that about another model next year and the year after that lmao. Anyway, it is impressive in that it's the first AI that I've actually seen tell a user if something it wrong in their prompt instead of blindly accepting everything as fact. Its still prone to hallucination though.

[balmy mist]

noghtwhisperrrrrrrrr ahhhhhh

[keen fulcrum]

What do you think?
https://www.together.ai/blog/deepcoder

[balmy mist]

my baby nightwhisper being unleashed on the world

[balmy mist]

is it fast?

[red sluice]

Google's dreamtides' kinda weak

[keen fulcrum]

Better than r1
(14b)

https://ollama.com/library/cogito

[north vale]

https://x.com/geminiapp/status/1909721519724339226?s=46

Google Gemini App (@GeminiApp) on X

📣 Deep Research is now powered by Gemini 2.5 Pro, our most intelligent AI model. ✨

This upgraded Deep Research is now even better at:

🔍 Finding & synthesizing information
📊 Providing more insightful reports
🧠 Analytical reasoning

Gemini Advanced users can access the new

X · Apr 8, 2025, 9:34 PM

[ocean vortex]

well no sht considering that's the model they started with lmao

Through a joint collaboration between the Agentica team and Together AI, we release DeepCoder-14B-Preview, a code reasoning model finetuned from Deepseek-R1-Distilled-Qwen-14B via distributed RL.

gains don't really look all that impressive considering this tbh

[torn mantle]

what

where

https://x.com/GeminiApp/status/1909721519724339226

Google Gemini App (@GeminiApp) on X

📣 Deep Research is now powered by Gemini 2.5 Pro, our most intelligent AI model. ✨

This upgraded Deep Research is now even better at:

🔍 Finding & synthesizing information
📊 Providing more insightful reports
🧠 Analytical reasoning

Gemini Advanced users can access the new

X · Apr 8, 2025, 9:34 PM

[keen beacon]

"Gemini Advanced users" joever

2.0 flash thinking

[north vale]

^

[brittle tiger]

Wow that's amazing. They are moving fast

[keen beacon]

best LLM for multiple choice questions?

gemini is good at math

needs to use its 2.5 version w/ deep thinking so the time for each question is long though

anything more accurate and faster that doesnt need deep thinking?

[north vale]

For most math gemini 2.5 is the best available right now at any price

If u have different tasks it might depend but 2.5 on avg is better than the rest

[balmy mist]

tmw gemini coder

[balmy mist]

yoo

[keen beacon]

yup already got gemini for math but for simple multiple choice

[balmy mist]

anyone tried the deep research yet?

[keen beacon]

heres a example

[balmy mist]

it has to be amazing

[keen beacon]

it really is

especially for coding

[balmy mist]

wait deep research is good for coding?

like you ask it to code?

[keen beacon]

yea i use it for LuaU

[balmy mist]

can you show me an example please

google just keeps winning lol

[keen beacon]

https://www.roblox.com/games/87936326400219/Tower-of-Rush-FREE-AURAS

[keen beacon]

not finished but made entirely with AI and some modules for optimization

datastore modules, etc

[balmy mist]

google made their stuff so abusable, i feel bad

like studio is free SOTA usage

then you have deep research

[keen beacon]

im a paying customer tho

xd

delete

dont expose method

they prob have lurkers in here

[balmy mist]

so i didnt make new ones

bet

[north vale]

multiple choice questions in what discipline? I'd expect there's not really a skill in llms that is "multiple choice questions", i'd think it's more about how knowledgeable they are in whatever topic you're asking questions about

[keen beacon]

yea but they might patch and add geolocation or sum type of tracking to ratelimit

gatekeep it

[balmy mist]

lmaoooo

fr

they got to

cause i feel bad at this point

[keen beacon]

im doing marketing but i need like a general one yk like chatgpt you can ask about accounting then also a micro economics question

idek how to explain it

prob just stick with gemini 2.5 and 2.0

[sage raptor]

with rocode ?

[north vale]

gemini 2.5 will prolly be best but there's a chance gpt4o is best if it's a very specific format bc it's usually better at following correct formats? but 2.5 seems the best bet

[north vale]

they ruined their impressive announcement by having the dumbest most shameful marketing ever

[torn mantle]

interesting

[north vale]

1. "At o3-min level" but if you scroll it's just o3-mini (Low)

It achieves an impressive 60.6% Pass@1 accuracy on LiveCodeBench (+8% improvement), matching the performance of o3-mini-2025-01-031 (Low) and o1-2024-12-17 with just 14B parameters.

2. [image]. NOT A JOKE.

[sage raptor]

wow

[north vale]

wait that's nuts

almost hard to believe? usually frontier models are like a few % apart in human preference

[sage raptor]

also openAi's deep research* is using o3 full

[balmy mist]

yoooo im in tears lmaoo

nahhh man what

im cryin yo

[sage raptor]

open ai is cooked

[balmy mist]

that is jokes

this why iyla left bro

dude was outty

[torn mantle]

i think google is on the right track

[balmy mist]

imma try deep research now, but i dont know what to search

[brittle tiger]

I think the word kids are saying these days for this is mogged

[keen beacon]

i just saw this

wtf

may have to subscribe

is anyone taking requests

🙏

"How Florida went from swing state to GOP stronghold"

[balmy mist]

i got you

nvm craig got it

but i got any others

cause i dont know what to search

deep research is just an agent at the end of the day

i wonder if we will get api for it, it technically is the best agent or way to search the web based on those benchmarks

like an agent for the web

this shows that 2.5 with tools is just on another level

imagine we get a 2.5 code cli like how we have claude code

bruhh, have the deep research built it and get other tools

i would pay for that, like a subscription easy

thats prob whats dropping tomorrow

this update to deep research was the first stage

tomorrow we get Gemini Code powered by 2.5 pro

and isnt google deep research faster than open ai lmaooo

damn

[north vale]

is this just a guess

[balmy mist]

nah its programmed at this point

trust

but yeah a guess lol

but high faith guess

why would they not?

2.5 cost less money to run(possibly) then claude 3.7 and 3.5 and anthronpic has a code cli

2.5 is better than 3.7 and has the google infrastructure behind it

[north vale]

that's probably just bc google has cheaper flop per dollar than anthropic

[balmy mist]

and they released the Deep research update today adn they said this gonna be a big week

[balmy mist]

exactly why we will see a Gemini Code, they gonna bury open ai and perplexity

i dont lol

thats only bc thats good with tools

but they are fixing 2.5 to work better with tools like cline, and other ides

its butt?

[north vale]

i doubt they will bury openai, and perplexity is irrelevant

is there a lmarena search leaderboard cooking? there's a search option to arena battles and search lb would be useful

[balmy mist]

bro openai is losing money and no longer have the SOTA model

[north vale]

perplexity value add is perplexing

[balmy mist]

i dont see how openai continues

especially when people already use google infrastructure, i could see if openai had the better model and it was cheaper or provided a better experience

but they dont

and google got gmail, youtube, maps, search, etc.. most normies not gonna wanna switch

[north vale]

4o has gotten much better with post training updates which shows they're getting good at post training. they are clearly very good at reasoning, google seems close but unclear if they are at the same level. openai is probably still at a similar level to google but probably releases their frontier models slower

so I just don't rly agree

[balmy mist]

but 99% of ppl heard of google

and they are integrating it into their systems

[north vale]

look at chatbot usage

chatgpt has 1000x name req of gemini

yeah they can add gemini to google tools

[lime coral]

Set system instruct

[balmy mist]

but chatgpt is not competing with gemini

they are competing with googel

[north vale]

but opensource sota will be useable for those things for day to day tasks by users in not long

[balmy mist]

do you use gmail?

[keen beacon]

any progress?

[north vale]

the sauce will be in the really intelligent, really expensive to run models, that the users will have little use for imo

[balmy mist]

what about maps? or youtube?

or drive?

you using drive or canva by chatgpt?

like come on lol

[brittle tiger]

OpenAI is incinerating money. Google has money printer and their sota model and deep research is 10x cheaper to run because of TPU and insane infra. If you think 600m chatgpt users will be sticky when Google gets agi I'll have whatever you're smoking

[lime coral]

Not true lol, at least with the imagen drama + don’t forget it’s on android. At least on my circle they know it

[north vale]

i just don't think the little products like "canvas" provide much value to these companies

or "gpt store" or whatever

[balmy mist]

you guys are sleeping on google

[lime coral]

Lot more request lol https://x.com/devsharma_8/status/1909728111744471097?s=46

Dev Sharma (@devsharma_8) on X

@GeminiApp is now amazing.
Best Model + Deep Research

20 Per Day!! Compared to @ChatGPTapp 10 per month.

For the same price. Crazy value. I find myself using it more.

Amazing work @AarushSelvan and @GoogleAI team.

X · Apr 8, 2025, 10:00 PM

[north vale]

nah google will blow up and become competitive but chatgpt is still very good

[balmy mist]

openai is fighting a losing battle, and dont got the pockets to truly compete and they losing people to other companies

[brittle tiger]

I firmly believe if timelines are early it's Google. If timelines aren't it will be a government

[balmy mist]

not really

[sage raptor]

" @ChatGPTapp
10 per month"

[balmy mist]

most of the stuff came from google

google just did not finish up what they started

i like openai pushing google

[thorny drum]

most people dont know the diff between chatgpt and AI

will be a long battle for gemini to take over market share but if their products are better its possible

[balmy mist]

most people dont know they been using AI for decades before the last 3 years

[north vale]

o3 has nuts scores, google hasn't come close to replicating those benchmarks yet

[balmy mist]

so it dont matter

its been integrated for years

now we are just integrating the SOTA into our systems

[north vale]

like both companies have strength it's a bit braindead reductive to just conclude "google has the sauce" or "openai is cooked" or some meme like that

[thorny drum]

thought 2.5 beat o3 on some benchmarks? anyways didnt o3 use like millions on compute to get those scores

[balmy mist]

they dont have to pay for gemini its free

and google gonna integrate with everything

youtube, gmail, drive, maps etc..

[north vale]

perplexity is a bubble company

[balmy mist]

no other companie has this reach

what??

you joking?

having gemini with maps has no value?

or drive?

or gmail?

[keen beacon]

been reading, good output

the tables are nice and i still find it crazy that ron desantis went from barely winning in 2018 (R+0.4) to the 2022 landslide

[balmy mist]

i do

[brittle tiger]

AGI won't have value

Lmao

[balmy mist]

and my parents do, my friends do, you are the minority

wait ai is already in those apps tho

do you not know that?

you can use ai for more than just a chatbot

optimization

integration

[brittle tiger]

Openai will never be profitable

[north vale]

yeah. it's fine not to compare o3 scores with billions of tokens generated vs gemini with thousands or millions. but an important thing about scaling test time compute is that companies that can squeeze out some intelligence out of spending more tokens will get ahead in terms of high quality reasoning. so probably they should start reporting stuff like cons@1024 or high numbers like that to show how well their models scale with more compute. if google scales bad with more compute compared to openai that might be a big openai advantage

[lime coral]

You fool

[north vale]

openai doesn't need to be profitable with chatbots, they will make plenty of $ if they can replace some jobs with ai

[brittle tiger]

Openai isn't sota

And they are expensive af

[thorny drum]

relax man they were sota as of like 3 weeks ago

[balmy mist]

im just saying google has the arms spread everywhere and it will be a easy transition for most people, openai only has their website and sdk and they are trying to branch out lol, claude has mcp and the cli and arr doing the saem

[thorny drum]

these things arent gonna matter in the big picture

[balmy mist]

but google has so much already

[brittle tiger]

Fine tuned Gemini 3.0 equivalent is designing tpuv7s right now. OpenAI doesn't have v1 of a chip. Just cash burning anime making homework helpr

[balmy mist]

the thing is, google can copy openai and anthropic, but they cant copy google

[thorny drum]

yeah the more access to compute + cash printer i think is a decent argument but these models are all pretty similar rn in the grand scheme of things

[balmy mist]

exactly

[thorny drum]

except llama lol

they're cooked

[brittle tiger]

-sent from backseat of a waymo btw

[balmy mist]

thats why the infrastructure that google has stands out, and what makes it worse is that google has the SOTA model rn and its free!!!

[north vale]

i also think llama is not doing well but i don't think they're cooked

behemoth spent around the same amt of compute as llama 3 405b! which is like a year old

which just shows they realized they overspent on 405b

and should focus on better scaling training than big rushed releases

[thorny drum]

we'll see maybe they gain some market share while they have the best model

[north vale]

and it will be 0-4 months behind SOTA

[thorny drum]

thing is i think ghibili did more for gaining market share than gemini 2.5 lol

[balmy mist]

that is true! but that can be copied their is no mota on intelligence or models

thats why google is in a good position

they are making it free

[north vale]

it did like 5 orders of magnitude more?

maybe 50 million users? and 2.5 got like a few tens - hundreds of thousands?

[brittle tiger]

I really only think what matters is whoever is closest to agi. That model rn is 2.5. recursive self improvement is gonna come and market share will mean nothing. It's winner take all

[north vale]

idk maybe i'm overestimating both

[balmy mist]

how can the other companies compete when the SOTA private models are free from google and then you have deepseek out here doing what they doing, openai, meta, and anthropic looking scary, idc who wins, but i dont see how the others can compete

true but there is no moat bro

[north vale]

maybe 20 mil vs 50k? would be 400x so 3 OoM

i'd be curious for data on that tbh

[balmy mist]

it dont matter what they have, it will be copied

[thorny drum]

i also dont think the cost of these models is really so bad. like i think most people atp are getting >$20/month of value from their favorite llm

[balmy mist]

true, thats a good argument

[north vale]

lol no they didn't

[brittle tiger]

Elon has been sandbagging them from get go. Doge uses grok and gemini

Trump's friend Elon musk who owns tesla

[north vale]

go outside bro

[brittle tiger]

Ppl think Elon isnt doing self interested stuff with his position. They are with every agency

[upper wolf]

It’s elon on an alt it’s gotta be him

[north vale]

he is but it's pretty minor overall, might be relevant later ig but probably minorly, and there's lots of reports of elon getting farther from trump's ear in the last few weeks

[brittle tiger]

I'm saying he's being corrupt and screwing altman

[north vale]

i just don't think the doge corruption is that much of a factor

trump does not seem biased against altman, he mostly seems to dismiss elon's beef with him

maybe there'd be an effect to the degree to which trump likes/dislikes bigco, which might affect the political benefits to google / microsoft / meta vs openai / xai / startups, but i'd expect that to be pretty minor

[azure minnow]

Who won

[north vale]

wtf

this isn't a battle

[barren prairie]

Google 😆🤌

[brittle tiger]

This from gemini 2.5 deep research on why google is going to win the race is very strong

https://docs.google.com/document/d/1u5OyQFZ4UsxY7OqomklvjRZzyCEFGzio84T0c0d_rEY/edit?usp=drivesdk

[leaden palm]

alt take

[north vale]

valuable take

[balmy mist]

https://huggingface.co/spaces/IjedMeer/clone-game

[brittle tiger]

It's personal beef. Just woke up one day? They had foresight to buy deepmind and start tpu project over a decade ago.

[leaden palm]

i don't think they did that because they had a vision for a world with agi

they did it because they could and it would be profitable, very unlike anthropic or openai's visions for the future

if you have a google agi manifesto please drop it in the chat

[brittle tiger]

Sundar, Larry and Sergey have been talking about agi and how it will be more important than invention of fire since before oai existed. I'm just saying that near thread is personal beef, which it is.

[upper wolf]

Imagine getting paid a full salary to not work lmfao

[torn mantle]

there is a term to this

a lot of companies do that

[raven void]

disagree with near

but it probably doesn't matter who wins, government will most likely control AGI

[alpine coral]

yeah gardening leave

it's pretty common, not just in tech (though 1yr is quite a long time)

[leaden palm]

which government do you expect to win / how do you expect them to use it

[ancient reef]

wait, did u just turn blue?

[leaden palm]

since yesterday

[ancient reef]

congrats 🎉

[leaden palm]

ty

[ancient reef]

np

[balmy mist]

anybody been playing with the deep research from google?

yall seen GSI Lab?

webdev and lmarena needs a display on ui to show when new models have been added, this discord should have an alert that we can check where we see the models that get added

[keen beacon]

where are the applications at 👀

[leaden palm]

i honestly don't know, i was just privately reached out to

[keen beacon]

ooh

[ivory schooner]

24k、spider不见的这几天，想它~

但愿Behemoth真的在24k的基础上构建😂

[harsh flume]

maverick

[harsh flume]

anonymous-test is def llama

[balmy mist]

studio new look

now you can compare models

night whisper is def coming

you can test the same model with different system prompts omgg

[torn mantle]

yea

[torn mantle]

looks much better

they also added 2 new models

[balmy mist]

i forgot you could always stream

[torn mantle]

gemini flash and veo

[balmy mist]

just never used it lmao

[balmy mist]

where

in studio or gemini app?

[torn mantle]

studio

it will added soon

[balmy mist]

the api

oh studi

wow

[torn mantle]

both

[balmy mist]

wtf

yoooooo

[torn mantle]

i dont think we will see nightwhisper this week tho

they cant just release all models at once

[balmy mist]

damn man, yeah they giving us to much heat that we dont deserve all of it

we been bad fr

SOTA video gen for free wtf and built into the platform i already use the most

you know what i think is happening

they are notcing that ppl are using studio the most

cause i dont even touch the gemini app, only for DR when i have the free trail and that was only like 3 times, and one time today to test it out

studio is very developer friendly

like out of all of the companies studio is my fav bc of the control i have

its almost like api access but still on app and free lmaoo

similar to openrouter and lmarena but my prompt almost never fail on studio

unless its a huge hughe prompt

[torn mantle]

yea thats crazy

im also glad they changed the UI

[balmy mist]

the one thing they need to add to studio is mcp and more tools that we can call and then its gg

me too

way cleaner

wait have you used the function calling at all in studio? i think i am underutilizing it @torn mantle

[torn mantle]

not yet tbh

[sturdy mica]

new ai studio ui

https://tenor.com/view/case-oh-caseoh-waffle-house-waffle-house-gif-10934642274965704175

[balmy mist]

didnt even know you could do this before lol

[sturdy mica]

not gonna bother reading message history cause i don’t care enough

[balmy mist]

wanna see if you can really hook up studio with like 10 tools to make powerful agents

[sturdy mica]

i’m pretty sure they default off in studio

unless i turned mine off and it saved or something

[balmy mist]

yeah it defaults to off

but funny they have that filter

[sturdy mica]

oh okay

yes

why even bother turning it on

[balmy mist]

imma really push studio to its limits this week

fr

[sturdy mica]

why would someone purposely enable restrictions on a model