general

[jovial sapphire]

*the requests

and then i created a tampermonkey script to intercept the checkpoint

and display it when it was on A/B

[ocean vortex]

I'm not sure they don't assign the same properties to different models...

[jovial sapphire]

they assigned the same

[ocean vortex]

in a/b

[jovial sapphire]

it wasnt random cuz

we all got same performance for same checkpoints

just go on google and type "x28 checkpoint gemini"

you'll see

i'm trying to get

an a/b

to show you

just got it

i'll show you how you can see the checkpoint

now

so first you wait until theyre done

[regal zealot]

you said you are a LLM researcher befoee.. and now you are a Security Resercher? So what? are you trying playing on us? OMG!

[ocean vortex]

Perhaps, I mean plausible... I miss when the days when you could just direct chat on lmarena some of the biggest unreleased models

[jovial sapphire]

check my twitter

when did i say i was

an llm researcher

lmao????

[regal zealot]

link?

[jovial sapphire]

i said i can create a group with a stanford ML researcher

[jovial sapphire]

yeah crazy

[regal zealot]

so thats im correct!!

[jovial sapphire]

i'm a security researcher

i can create a group with someone that is an llm/ml researcher

[regal zealot]

LLM same as ML.. but LLM is bigger..

[jovial sapphire]

i didn't say i was

[regal zealot]

oh my mistake..

[jovial sapphire]

np !

[regal zealot]

so what you can do with your programs?

[jovial sapphire]

hahahah oh my god

the result is amazingh!

prompt: "
User
Create a neo brutalist web page, esthethic that looks like a modern museum., 1000 lines, multiple pages, effects etc."

one shot

@ocean vortex

[regal zealot]

looks like a site..

[regal zealot]

he might got to sleep..

[jovial sapphire]

yes it is a very beautiful one for a one shot prompt!!!

SOTA in coding

@ocean vortex

so when you submit A/B test you go here

[regal zealot]

this is a one shot?????!

WoW! then if it was one shot!!

i couldnt do that inone shot!!..

[jovial sapphire]

yes it is one shot

and then you can go to ```"{"temperature":1,"topP":0.95,"topK":64,"jl":[],"maxOutputTokens":65536,"safetySettings":[[null,null,7,5],[null,null,8,5],[null,null,9,5],[null,null,10,5]],"enableCodeExecution":false,"enableBrowseAsATool":false,"Fj":false,"responseModalities":[],"zf":false,"enableSearchAsATool":true,"googleSearch":[],"thinkingBudget":-1,"outputResolution":"1K","model":"52c1a3b8c57b46fb"}"

and at the end you have ""model":"52c1a3b8c57b46fb"

and this is the checkpoint

[ocean vortex]

So where does it say "x28..."?

or the actual checkpoint this is

[jovial sapphire]

"model:"

so when it was x28

it used to say

"model:x28blablabla"

but now it's gone

[regal zealot]

you mean its now renamed to 52blablablabla?

[jovial sapphire]

html code of the website it made

[jovial sapphire]

no, it's a different checkpoint

but very good

it's gemini 3

[regal zealot]

oh okay..

[jovial sapphire]

this is gemini 3 result

no other models

can one shot

this @ocean vortex

that's why i'm saying it will be sota

[regal zealot]

and how did you connect to gemini 3.0? before it have been released ?

how can i?

[jovial sapphire]

i know it is

youll see

[regal zealot]

how can i?

[jovial sapphire]

make urself an automating script

[ocean vortex]

that seems encoded

or just obfuscated

[jovial sapphire]

no it's not

broooo

fdsqgopàg^ksdqfkdsqf

[ocean vortex]

?

it literally is lol

[jovial sapphire]

bro

when you do another

a/b

it's not the same

and sometimes it comes back

and when it does

it's the same performance

that you had previously

with the same checkpoint

is your head full of air

i'm done with this debate

[regal zealot]

lookes like u typed very fast and got removed very fast..

[jovial sapphire]

you can't understand

just nevermind blud

[regal zealot]

cani connect to gemini 3.0?

[ocean vortex]

You are not understanding what I'm saying. It is clearly obfuscated as that is not the real name. Not saying that you can't get the same one the 2nd time

[tiny halo]

hey guts does anyone know why i cant send photo to ai chat it wont let me send

[ocean vortex]

are you ok?

lmao

[regal zealot]

please.. can i chat to you, from me to you?

[gaunt spade]

how did u get a gemini 3 checkpoint on AI studio

[ocean vortex]

Let him do his 'security research' with inspect element Network tab 🤓

[regal zealot]

😆

[regal zealot]

lmaaooo!!!

[gaunt spade]

@jovial sapphire yo bro how do u get A/B tests on AI studio

cuz i never got them

is it locked to some regions?

[regal zealot]

maybe

[gaunt spade]

i wanna try Gemini 3 checkpoints on AI studio

maybe they're better than riftrunner

in coding

[regal zealot]

.

[rustic lichen]

pls could u add an option to share chat?

[regal zealot]

u mean in lmarena?

[jovial sapphire]

bro

i know

it's not the real name

[rustic lichen]

yh

[jovial sapphire]

captain obvious

thats why it's called checkpoints

and we call them by their codename (x28, ECPT)

[jovial sapphire]

try usa

[gaunt spade]

how can i encounter the checkpoints

[jovial sapphire]

usa

ip

i can run prompts for you if you want

if you have very good prompts

[gaunt spade]

does it require some prompting to get a checkpoint

[regal zealot]

its cool if from the site u could even download the chat to ur device.. beacuse i want that !

[jovial sapphire]

nope

0

[rustic lichen]

yh

[gaunt spade]

hmm lemme search through my prompts

[jovial sapphire]

coding

prompts if possible

[gaunt spade]

yeah its coding, a 3d simulation

wait imma look it up

[regal zealot]

is your chat in lmarena is direct --chat?

[rustic lichen]

yes

[gaunt spade]

also are the new checkpoints better/on par with x28/x58

[jovial sapphire]

some of them are on par

but x28 was very good overall

[gaunt spade]

i love to see non-nerfed outputs

[regal zealot]

then we must develope together a tool that gets the chat messages very easily..

[gaunt spade]

from gemini 3

[jovial sapphire]

yes x28 was crazy

well if you have other coding prompts

don't hesitate

[gaunt spade]

yeah i saw crazy demos

[jovial sapphire]

i'll run them now

[gaunt spade]

made with x28

wait imma think

of my prompt

[jovial sapphire]

youre the first one who knows what checkpoints here are lol

happy to see not everyone is a skeptic

[regal zealot]

how can i join the x28?

[jovial sapphire]

you cant lol

its gone

now

[gaunt spade]

well, i've known checkpoints since lithiumflow

lol

like 3-4 weeks ago

[jovial sapphire]

yup, a/b testing first starts on

ai studio

then lmarena

[gaunt spade]

yea lithiumflow was good too

but x28 is the best

[jovial sapphire]

not as good as

x28 yes

but still good

any coding idea? could be anything

[gaunt spade]

'''Create an interactive 3D wormhole simulation using three.js, featuring a tunnel-like geometry with animated textures that give the illusion of depth and motion. Implement custom shaders or materials to simulate swirling energy, stars, or nebulae moving along the tunnel, and animate the camera to travel smoothly through the wormhole. Include controls to adjust parameters such as tunnel length, speed, color scheme, and distortion intensity in real time. Provide well-commented JavaScript code, an HTML setup snippet, and brief explanations of key three.js concepts used (geometry, materials, shaders, animation loop, and camera controls).'''

thats a good prompt

[jovial sapphire]

Okay!

ty

let's hope i get a good checjpoint

[gaunt spade]

no problem

[gaunt spade]

how many checkpoints are live rn on AI studio

and how much time does it take you to encounter one?

js curious

[jovial sapphire]

i've seen 3-4

some are very bad

[jovial sapphire]

sometimes a few mins

sometiomes 30m

[gaunt spade]

maybe they're flash lite

lol

[jovial sapphire]

here

are u on pc?

[gaunt spade]

damn

im on my phone rn

[jovial sapphire]

shi wait i can upload

[gaunt spade]

did it come out already

[jovial sapphire]

the code to codepen

[jovial sapphire]

nope not yet

but

i can show you one run i did earlier

[tiny halo]

hey guys does anyone know why i cant send photo to ai chat it wont let me send

[gaunt spade]

oh

i would like to see it

[jovial sapphire]

prompt was to create neobrutalist website

[gaunt spade]

thats a good prompting idea tbh

was the output any good

wait brb

i got something rn

[jovial sapphire]

codepen is down...

[jovial sapphire]

yes

here

https://bold-power-hamh.pagedrop.io/

one shot

"Create a neo brutalist web page, esthethic that looks like a modern museum., 1000 lines, multiple pages, effects etc."" prompt

[gaunt spade]

it nailed your prompt tbh

was it the best checkpoint's output

[jovial sapphire]

it is one of the bests i have tried rn yes

[gaunt spade]

im pretty sure, there are other gemini 3 models in there like flash and flash-lite

thats why some checkpoints suck at your attempts

[jovial sapphire]

yes

surely

yay!

your prompt got an ab

let's see

[gaunt spade]

wow it didnt take that long

wait do u use a bot that spams prompts

[jovial sapphire]

this is so cool hahaha

[jovial sapphire]

yes

[gaunt spade]

hmm, do u mind sharing it

[gaunt spade]

wait paste the html file here

[jovial sapphire]

i use some api keys inside

i cant rn

[gaunt spade]

only 10KB, lets see

[jovial sapphire]

this is very cool

imma try with gemini 2.5 pro

[gaunt spade]

how about riftrunner

its a gemini 3 checkpoint too

on lmarena

lets see how it compares

[jovial sapphire]

same checkpoint

as the one who did

the website

[jovial sapphire]

did this

52c1a3b8c57b46fb

[gaunt spade]

how do u know

[jovial sapphire]

checked

[gaunt spade]

maybe riftrunner is a little worse/better

[jovial sapphire]

i'm talking about

my website

neobrutalist

[gaunt spade]

oh bruh

[jovial sapphire]

.

[gaunt spade]

i thought u meant that the checkpoint is riftrunner

[jovial sapphire]

imma try the same prompt with g2.5

oh no nono

its far better imo

i'll try SVG

[gaunt spade]

imma try it with riftrunner, maybe it produces something interesting or different

[jovial sapphire]

trying with classical g2.5 pro

super trippy

(gemini 3, not gemini 2.5)

LMAO

gemini 2.5 pro completely failed @gaunt spade

doesnt even work

[balmy mist]

you can still get gemini 3 through a/b on studio?

[jovial sapphire]

yes

FPS is bad

[gaunt spade]

which one is that

[jovial sapphire]

wait

[jovial sapphire]

same as the neo brutalist one

i sent you the html

[gaunt spade]

oh yea

i thought u tested another one

im yet to get riftrunner in lmarena

its quite difficult

[gaunt spade]

u can, but i think you need to be in the US now

[jovial sapphire]

no other model get it

trust me

[regal zealot]

have u ever jail breaked gemini ? grok? , i have and worked with them.. , but the next day , it have been restricted.. even if i tried encoding the prompt.. looks like lmarena shares the chats with companies , and maybe have told them i had tried how to make something , if i told him as a test to the jailbreak prompt , he will say , "Sorry, I can't assist with that."

[gaunt spade]

well riftrunner might be close, cuz its Gemini 3 too

havent u tried it yet?

its a little worse than lithiumflow

i guess

[jovial sapphire]

i found it very bad

[regal zealot]

i guess even gemini 2.5 pro , got updated everyday...

beacuse it got better by everyday...

[gaunt spade]

wait

can u try another prompt for me

[jovial sapphire]

yes

trying opus 4.1 on the wormhole thing

to see if it performs better than g3

[regal zealot]

so u mean if i tried to edit the thing in browser, i can connect with gemini 3.0?

[jovial sapphire]

it's nothing to edit

it's just a/b

just wait until it releases bro

it will release in a few days

no worries

[gaunt spade]

can you do this one "Create a fully functional, identical and playable DOOM (1993) clone created in three.js <{5000 lines of code}>"

[jovial sapphire]

it crashes

after

3k lines average

[gaunt spade]

oh

make it 2000

[jovial sapphire]

cuz of ai studio limits

okay!

thanks

[regal zealot]

how can i understand what does a/b means? u mean a dual ai of gemini 2.5-pro making them selfes a 3.0?

[jovial sapphire]

launching now

[gaunt spade]

np

[jovial sapphire]

nonono

[jovial sapphire]

can you explain this guy

[regal zealot]

a/b isnt means a for gemini and b for other one? working together?

[jovial sapphire]

basically on ai studio

when you send more messages

you have a probability

that they will ask you to choose between two answers

for testing purposes

and one of them is gemini 3

[regal zealot]

i rememberd that..

[jovial sapphire]

launched your test buni

[regal zealot]

howw did u know??

[gaunt spade]

a/b is a rare testing thing that comes up in Google AI Studio sometimes, where you can choose between two models by the one who responds better, choosing A or B

[jovial sapphire]

looks very bad imo @gaunt spade

this is Opus 4.1

for wormhole

looool

very cheap looking particles animation

looks overengineered

and doesnt even look like a wormhole

[regal zealot]

i have rememberd that happening alots of between a few long period of times..

[jovial sapphire]

doesnt look like a wormhole at all

gemini 3 one was so much better

[regal zealot]

howw did u know if it was really a 3.0???

and how can i connect to gemini-3.0-pro?

[gaunt spade]

someone made that prompt in X28, i remember how crazy it looked. but it still looks good on the new Gemini 3 checkpoints

[regal zealot]

@jovial sapphire

[gaunt spade]

because the demo looks very good

and its detailed with proper coding

[regal zealot]

and how can i connect to him?

[gaunt spade]

you need to be in a specific region (the united states)

and run some prompts on AI studio until you get the A/B testing checkpoint

that means a better model will produce your output

and you get to choose between one of them

[jovial sapphire]

it takes a long time

[regal zealot]

you are same as gemini and ill tell you same as what im gonna tell him when he does that:

" Why Are You making the things more complex and trying hiding the true solution to me? you could easily told me use v.pn!! please dont do this again ever, as i have eyes, and can know , so u better dont hide, otherwise ill use gpt "

[regal zealot]

and how can i know which one?

[jovial sapphire]

u try both

and u see whos better

you have choice between both

[regal zealot]

so i can use both ? instead of seleceting one? and how?

[gaunt spade]

@jovial sapphire yooo

i finally got riftrunner

on lmarena

svg test

[jovial sapphire]

lets go

[gaunt spade]

it does so well

[jovial sapphire]

yes but

you shouldve tried

wormhole 🙁

[gaunt spade]

dont worry

cuz i

[jovial sapphire]

okay try again

[gaunt spade]

didnt vote lol

[jovial sapphire]

nice!

[gaunt spade]

i know how riftrunner does

and i dont vote

if it got it

so i get to chat with it

a little more

@jovial sapphire imma make the wormhole and then doom

[regal zealot]

looks like his feet is connected with the cycle.. 😆

[jovial sapphire]

okay!

waiting on doom

no a/b for now

[gaunt spade]

bruh riftrunner FLOPPED SO HARD

[jovial sapphire]

LMAOOO

wth is that

bro

[gaunt spade]

i voted and it turned out its indeed riftrunner

[jovial sapphire]

thats what im saying

people here dont believe me!!!

[gaunt spade]

wow i didnt expect it to fail at the wormhole

[jovial sapphire]

AI studio is the real gemini 3 pro

[regal zealot]

LMAOOOOOO!!

[gaunt spade]

riftrunner is a nerfed gemini 3

sadly

[jovial sapphire]

if they release this one

i will be so so so disappointed

[regal zealot]

the code looks like from 2016!!! 🤣🤣🤣🤣

[gaunt spade]

well well well, this checkpoint is the same on that appeared on canvas on gemini app 😭

the outputs were the exact same

[regal zealot]

i didnt laugh this hard on whole this day until nowww!!!!!

[jovial sapphire]

no canvas

was not as good

as ai studio one

[gaunt spade]

yea

[jovial sapphire]

thats why

[gaunt spade]

it was like riftrunner

[jovial sapphire]

im hella confused

[gaunt spade]

thats my point

[jovial sapphire]

ok mb

what will they release

id be so so sad

[gaunt spade]

they released the worst checkpoint on canvas

thats my point 😭

[jovial sapphire]

yea....

i dont know what they will release

in the end i mean

[gaunt spade]

maybe riftrunner is gemini 3 flash

with some thinking

cuz rift-runner sounds like a FAST model

for its codename

[regal zealot]

@hallow hemlock maybe u tried the Gemini-3.0-ultra-lite 🤣🤣🤣🤣

[gaunt spade]

did u get a checkpoint

on my prompt

[jovial sapphire]

nah crazy

over 100 runs

no ab

[gaunt spade]

also, do u get rate limits

on AI studio

[jovial sapphire]

lets go!

yes i do, but change google accounts

and its gone

[gaunt spade]

YEA

[gaunt spade]

wait what

how many accs u got

[jovial sapphire]

+100

i buy phone numbers

for a few cents

online

[gaunt spade]

wow

[jovial sapphire]

costs me like

1$ for 50 numbers

[gaunt spade]

i think theres a way to bypass the phone number thing

but maybe it got patched or something

[gaunt spade]

US numbers?

or what country

[jovial sapphire]

any country

google takes all

first result not crazy

[gaunt spade]

are there enemies

and does ur hp go down

[jovial sapphire]

yes

wait

actually no it s kind of good

enemies walk towards me and attack me

when i approach them

[gaunt spade]

which type

[jovial sapphire]

idk it's one type

it did anotherone

[gaunt spade]

the red eye thing?

[jovial sapphire]

but one error

ill fix it

[gaunt spade]

this actually looks close to my riftrunner doom

[jovial sapphire]

here try it

i'll see if it's the good checkpoint or not

[gaunt spade]

https://123.nekoweb.org/ai/stealth/riftoom.html

heres my doom made by riftrunner

[gaunt spade]

pretty good

[jovial sapphire]

yup

i'll fix the one that has

one error

and see if its better

[gaunt spade]

lets see

[regal zealot]

looks like there is just one monster and there is no good controls..

[jovial sapphire]

i think the one we got

wazs bad checkpoint

actually no it was the same

52c1a3b8c57b46fb

here check this

[jovial sapphire]

i think problem here was a very bad prompt, it was very short

"
User
can you do this one "Create a fully functional, identical and playable DOOM (1993) clone created in three.js, all in a single html file, 2000 lines of code."'"

still better than every other model tho

[jovial sapphire]

any other

prompts?

[jade egret]

bro chatGPT

you obsess with chatGPT?

[knotty fable]

https://youtu.be/NBZv0_MImIY

[gaunt spade]

damn

good video

[jovial sapphire]

yo

any prompts?

[regal zealot]

i have....

[regal zealot]

but its not a web or for web.. yes it csn be a web but its a fornt end could be a web..

nevermind..

[jovial sapphire]

i need web

[regal zealot]

how about explaining monero how it works in web page for a teenager like me cannot stand how monero works ?

like i know bitcoin how it works.. but monero i cannot stand it ..

like i cannot stand how it works in hiding the amount of current of what the sender have and to the reciever and everything is hidden.. i cannot stand that..

[jovial sapphire]

ok

any style?

[regal zealot]

everytime i asks ai about it such as gemini 2.5 pro , it cannot make me stand it..

[regal zealot]

yes , just tell him make it better and standable for a teenage guy such like me..

[jovial sapphire]

okay

will try now

[regal zealot]

just dont host it ur self online.. just give here the .html and ill host it myself in the terminal..

[jovial sapphire]

"Your task is to create a single-file, highly interactive HTML educational experience designed for a teenager who understands Bitcoin but is confused by Monero's complex privacy features. Focus on using relatable, casual language, and high-contrast, brutalist styling to demystify how Monero hides the sender, receiver, and transaction amount. The entire page must feature accurate, step-by-step animations or interactive visual analogies that explain the core mechanics of Ring Signatures, Stealth Addresses, and Ring Confidential Transactions (RingCT). The final product should be a visually striking and functionally clear demonstration that makes the user finally "get" the system. A neo-brutalist style axed on Monero visual identity. Make a masterpiece, minimum 1000 lines."

this is the prompt

[regal zealot]

u know how did u make the prompt?

[jovial sapphire]

asked gemini

i know monero

[regal zealot]

also do u know that i dont have gemini 3.0?

[jovial sapphire]

i use it

[jovial sapphire]

yes ikr

no one

[regal zealot]

alright.. ill try gemini 2.5 pro..

[jovial sapphire]

i will try Gemini 3

and send you the html

[regal zealot]

alright..

if he did make a short one.. say to him that the user dont love the short answers and wants a longer ones..

and tell him he will read it..

also tell him the user know what the pgp is.. and what the signature meaning and the publickey means..

[jovial sapphire]

Hahaha so cool

this is crazy

@gaunt spade

[regal zealot]

why u didnt give the html file or code?

cuz i cannot interact with it..

such as in the image , i cannot press buttons..

[regal zealot]

thats funny what he is saying at the end.. dont do something.. HahHaha

[jovial sapphire]

this one is a different checkpoint "ovdf5tin1bp9d2o3"

@regal zealot

one shot is crazy

[regal zealot]

i still laugh at the end of it..👀

[jovial sapphire]

hahah ayes

but you see, the result is amazing

gemini 3 is very good

[regal zealot]

yeah.. it could make it better such as like of what i imagine of how the sender or receiever when it wants to send , i thinking of maybe if the node of blockcha1.n it uses if the hidden hands knows it.. perhaps they might know him..

or may thought is wrong?

let it fix it.. so i can know more about moner0.. and also know why my only-view wallet needs alots of time just for syncing..

[jovial sapphire]

it explainedf

everything

[regal zealot]

so u mean my thought is right?

[burnt sinew]

https://tenor.com/view/xmr-monero-privacy-getmonero-end-the-fed-gif-20646141

[jovial sapphire]

ask gemini 2.5

[regal zealot]

as of what i see?

[jovial sapphire]

ask your questions

to ai

[regal zealot]

delete this NOOOWWWW!!!

[regal zealot]

you have been blocked..

[regal zealot]

alright after i did blcok , now im comfertable, that all i see is a "blocked message--Show"

[regal zealot]

alright.. so sorry..

[jovial sapphire]

no its ok bro

dont apologize

its an advice

[regal zealot]

i rememberd that i asked that question in the old past of time when the 2.5 was new..

but it didnt yet explained to me and maked me stands it its secure , so its not have that issue that i thought of..

[jovial sapphire]

you ask bad questions

thats why u dont get good answers

whats ur language?

[regal zealot]

how and why its bad?

[jovial sapphire]

i meant if u dont get good answers

maybe its because you dont ask the right questions

[balmy mist]

G3 out?

[jovial sapphire]

not yet

this week yes

[jovial sapphire]

but i can access

via ai studio

if u have good prompts

i can try for you

[tiny halo]

can someone help me please i have an issue

[regal zealot]

huh? , u mean my question is bad or my language? , i mean when i did ask it, i did in my own language witha little english..

[jovial sapphire]

what

is your main language?

[jovial sapphire]

.

[tiny halo]

i cant import photo to ai chat why please help

[balmy mist]

Yeah let me prep them lol

[jovial sapphire]

oktysm

[balmy mist]

Which one?

[tiny halo]

claude 4-5

[regal zealot]

if u have some respect to my privacy, u better not have asked me here..

[balmy mist]

On which app lmarena?

[tiny halo]

yes

[jovial sapphire]

bro

its not private

who cares

[regal zealot]

fe.ds may cares..

[balmy mist]

@echo aurora Can u help with this?

[tiny halo]

u dont have my issue ?

[regal zealot]

even if i didnt do anything wrong.. they still logs me in their logs..

[balmy mist]

Could be an issue with the app, but send screenshot of error message to help them debug I’m not dev but I can guide u in right direction lol

[echo aurora]

Sry what's the TLDR? I was headed out the door, how can I help?

[balmy mist]

Wassup bro, how u been?

[tiny halo]

hello please help me i have an issue i cant import photo to ai claude sonet 4-5

[balmy mist]

Paste error too as an image if you can

[tiny halo]

ok

[echo aurora]

Can you fill out a post in #1343291835845578853 if you haven't already? I won't be able to answer now, but that's the best way to flag and technical issues.

[balmy mist]

Bro same

[balmy mist]

Thanks! Next time I’ll just forward them them lol

[tiny halo]

heere

[balmy mist]

But u got this tho, u using ai to help?

[tiny halo]

i did but noone answer me

@balmy mist its jpg

[regal zealot]

brah whats even that slang ?? "ADDH"??

[craggy night]

How to generate image to image please

[balmy mist]

Try refreshing(ctrl + shift + r), I’m outside right now, but I could help when I get back home,

[echo aurora]

Ah yes, this model IIRC we had isssues with enabling vision with. At the moment it's not going to work with image upload sorry to say.

[regal zealot]

maybe "AD OF UHD?"

[tiny halo]

wich ai model has import photo

[tiny halo]

i did stil not wokrs

[balmy mist]

I’m the same way, I’m tryna start my family now, that la how far ai pushed me away from tech and I minored in it lol

[balmy mist]

Hmm codex

Gpt 5

But 4.5 should work

[tiny halo]

wait actualy working on gpt but all model claudes has this issue

[balmy mist]

That’s weird, there could be an issue with Claude models in arena, dm I got something u can use

lol yeah 23

Finish college a while back

But kinda miss it, adulting hard bro😂

Man i would say it get better but … lol im thinking about changing careers

Im a software engineer rn, but I want to branch out so bad, i hate being at my desk at home all day, ik its a privilege but this making me crqzy lol

[gaunt spade]

what grade u in

[gaunt spade]

just be chill and enjoy some moments on different stuff

u dont have to always learn about AI

[molten cypress]

hello here

i need some help

after uploading my image for video animation and prompt what key to press for generation because enter isn't working?

[balmy mist]

U right, it’s just so hard not to cause it’s become everything now tbh

[sharp mirage]

Hi @echo aurora Is there any update for the UI?

[jade egret]

lmao chill why u suddenly mad

[barren pine]

Hello, I want to learn, share ideas, and connect with others passionate about AI.

[crimson sky]

hello, want to learn prompts, generate videos and share ideas

[latent jungle]

Proof of your success?

[ashen plaza]

<@&1349916362595635286>

[ashen plaza]

its a scam bot

[latent jungle]

As I thought

[cloud zinc]

$100k wow

[whole swallow]

What it was?

[hardy lion]

there was a spammer who said you can make 100k by doing whatever he said or some nonsense

[fleet lintel]

scammers everywhere... i hate them

[whole swallow]

https://tenor.com/view/diddy-gif-8961692530157879891

[jovial willow]

Does google login not work for anyone else

[round sedge]

its still in every other channel

[sudden coral]

i cant even open lmarena cf under attack just spins forever no matter what browser i try

[patent bane]

same

[inland violet]

guyz du u beliv mi

[rocky mauve]

Best model currently for coding python tasks?

Don’t want one with limits though

[keen beacon]

my feet hurts

dang

[real rampart]

Hey everyone! I've been sifting through a lot of posts here trying to find information about two battle arena models: Crystal and Quasarflux, and I've only found guesses that Crystal is LLama and Quasarflux is Grok. Is there any definitive information published anywhere, or is it always a secret? 👀

[zealous sparrow]

New models?

@quartz light are these models new

[real rampart]

I saw Crystal a long time ago, about a few months ago, and Quasarflux, yes, it’s new.

[hollow imp]

That's chatgpt

@ocean vortex bro even your system prompt is not doing anything against the heavily nerfed gpt 5 pro

The custom instructions are not un nerfing it

[ocean vortex]

The prompt you are trying it with nearly all current models are gonna struggle by chance depending on their tokenizer. As far as I'm aware nothing major has changed with their current tokenizer. So essentially the model can't get this right without being overfit. You could test the tokenizer and how it performs or to what extent reasoning can remedy it, but then you would need much more prompts like this than just 1 random word. Different models are gonna fail on different ones

[hollow imp]

@ocean vortex you said your prompt will make it think longer, is 4-6 minutes long? I tried many tasks which should take long stem and non stem

[ocean vortex]

yeah 4-6min is not bad. I'm kinda fascinated by the 5.1-chat performance though currently. It is impressive how good it performs for how fast it is

But it's hard to say definitively if system prompt is really the only change their website to API

on API it is much more blunt

Reminds me of grok4 while being as fast as non-reasoning model, on API

[hollow imp]

Bro before it took 10-12 minutes for lower tasks than what i tested

And the quality is worse

And I'm not talking about 5.1 I'm talking about 5 pro

I mean how can gpt 5.1 heavy thinking take longer than 5 pro and give better answers

[ocean vortex]

For pro there's no 5.1 Pro yet afaik

[hollow imp]

Wtf is that nerf

It was never close to 5 thinking heavy

[ocean vortex]

on API I only see 5.0 Pro

[hollow imp]

@ocean vortex bro what do I do

😭

[ocean vortex]

I don't have ChatGPT Pro. It's a waste of money as far as I'm concerned lol

[hollow imp]

I don't too

[ocean vortex]

but hence can't comment on it much

[hollow imp]

I'm using business subscription free trial

[ocean vortex]

Do you have any prompt that you ran earlier and tried now without any changes?

Cause it''s very hard to know without retrying the same exact thing

You may think it was 'easier task', but models do not exactly always work the way you may expect them to. They will end up thinking longer for tasks that you anticipated to take less.

And then the quality is next to impossible to compare if you are comparing different tasks

[hollow imp]

Yes

Many infact

[hollow imp]

See this
https://www.reddit.com/r/ChatGPTPro/s/NS03WRx077

[quartz light]

yeah but i havent seen crystal

btw look

[keen beacon]

guys i have a question

[quartz light]

my thinking injector for any model
it worked on grok's stealth model

[keen beacon]

is chat gpt 5.1 familiar to chat gpt 4o

or im tripping

[ocean vortex]

Seems like there's a lot of the usual reddit shtposting there to be completely honest. I don't see a single prompt and output in there.

[keen beacon]

anybody bro

i need to know

[hollow imp]

What is shitposting?

[hollow imp]

Yes

[keen beacon]

its similar

fr?

in chatgpt?

[hollow imp]

Yes

Frfr

[ocean vortex]

Comments not based on reality but rather unfounded emotions, in this context

[keen beacon]

on god?

say on god @hollow imp

because chat gpt 4o is like a human buddy in old days

[ocean vortex]

Can you try one now, to compare it directly all things being equal? Also shorter reasoning alone does not mean it is worse. It is actually a big plus if the final answer turns out better.

[hollow imp]

THE FINAL ANSWER TURNS OUT SIGNIFICANTLY WORSE

[keen beacon]

@hollow imp yo say on god

rn

bro

I NEED ANSWERS!

[quartz light]

bruh

just try it yourself

[hollow imp]

On skibidi

[ocean vortex]

Then show one example of it like I asked lol

[quartz light]

🤯

[keen beacon]

say on god bro.

[quartz light]

on skibidi is everything

[keen beacon]

if you say it, i believe...

[hollow imp]

I can't, that prompt has some personal stuff

[hollow imp]

On taro sakamoto

[ocean vortex]

Cause that reddit post does not document a single one as far as I can see

[keen beacon]

imma search skibidi, wait

[ocean vortex]

just a pointless thread for the most part

[keen beacon]

it ain't everything

but i saw toilet

[hollow imp]

Dom please give some nudge phrase to make gpt 5 pro answers more quality

[ocean vortex]

you said there are many. Pick ONE you can share. 🗿

[stray aspen]

lmao

[quartz light]

yall

is ts prompt too much

[keen beacon]

bro hell no way

i have a limit on go chat

in chatgpt

thats tuff broooo

[wheat onyx]

As good as 5.1 is, very excited for gem3

[zealous sparrow]

5.1 sucks at coding wym

[wheat onyx]

Did I say I'm coding?

No idea how good it is at coding, but clearly it's the only use case...

[stray aspen]

gemini 3 is close

[wheat onyx]

This week... I will be using it a lot

Will be curious how flash is compared to 5.1 too

Oai really neutered 5.1 by disabling legal assistance too. So gemini will be an even bigger deal there.. It was already better than many lawyers

[wheat onyx]

Yeah they told gpt not to do any legal stuff. It was already better than a lot of lawyers I'm familiar with in gpt5.0

So disappointing. I know people will be switching to gemini.

If you're a lawyer that's not double-checking the legal reasoning, or are using any case law, you deserve to get screwed

[zealous sparrow]

No, whats your main use case?

[neat apex]

I rather be judged by gpt 5 than by an average people honestly immo

So gpt 5.1 sucks?

[sour spear]

No, it's simply not better than 5

[wheat onyx]

A few.

Translations, creative writing, legal, summarizing, transcriptions of audio. Probably more I'm forgetting

[wheat onyx]

No it is better than 5...but incrementally. And then they disabled legal assistance

[neat apex]

I mean that

Gpt 5.1 is clearly way better at all

[wheat onyx]

Too many idiot lawyers used gpt without using their brain, so oai disabled it... Disable the lawyers instead

[neat apex]

Yeah, that is a issue

Gpt 5 is better than a average lazy lawyer, but he will be barely feed of context

[wispy quarry]

Hi, I'm new here 😄

[wheat onyx]

Yup. If you go through it many times it usually gets there (with better than avg reasoning).. Rarely on first couple attempts

for translations and transcriptions, AI is pretty much good enough (as long as you give it just enough context to know what it's about)

[keen beacon]

imma use deepseek ai bro

[warm robin]

Gemini

[wheat onyx]

GPT 5.1medium is around 72% (plus), and high is around 76% (pro)

is SWE verified not the best benchmark for coding?

[regal cargo]

anyone know why its not working?

[ocean vortex]

🔥

[ocean vortex]

It isn't. It's only part of the story and a coding benchmark that is being exploited the most currently

And is being exploited for this very reason. People like yourself thinking it is the best 👀

[regal zealot]

lmao.. isnt that a bubble..

[ocean vortex]

Anthropic did it first, now Google and OpenAI doing the same. Focusing on SWE much more than they did before. Things like codeforces, LCB, Aider etc are secondary now

Kinda predictable if you ask me. They are always gonna focus on the things people are looking for. SWE is also easier to rig than having to worry about all the alternatives equally, in my opinion. Like with o3 they had to test a ton of different coding metrics. For 5.1 people are kinda happy just with improved SWE... 👀

[wheat onyx]

i dont code, i dont know either way. just curious. Are there better benchmarks?

[neat apex]

I also keep using deepseek xd, 3.2 is goated

[neat apex]

Maybe, but now it will be benchmaxxed even at unwill

[ocean vortex]

It's not a bad metric, but there isn't really 1 definitive coding benchmark that would stand in isolation of everything else. Variety is always important against contamination. So the score is only 'valid' if the scores are high in other related coding benchmarks as well

Not correct on what exactly?

[neat apex]

It happens mostly because they try to adjust it "better"

[keen beacon]

3.2?

3.2 good?

[jade egret]

yea

[neat apex]

Just like, lmarena nerfs Gemini 2.5 pro somehow since it is weird to it be better than sonnet 4.5

[ocean vortex]

the order varies a lot depending on what you choose to look at. For LCB just about all Claude models suck but this is still coding

[neat apex]

It is only partially true, for example an model that have higher world knowledge (Gemini 2.5) can outperform sometimes other model like o4 mini, not because the quality of model is consistent, but because the sum of competences

[ocean vortex]

You are way oversimplifying it. There are many models from each of them. And I'm still not sure what you didn't agree with in my initial message you quoted tbh

yes

[neat apex]

4o is bad at coding, but he knows ball and perfoms well in more standardrized tests

[ocean vortex]

incorrect

lmao

[neat apex]

I dont see a reason to not count human judged tests

And also it keeps being true in automatized or LLM judged tests

Curious, i wanna read that

How i said, i am being objective and not affirmed they are really good

[ocean vortex]

How do you think current benchmarks were made if they aren't human tests? By aliens?

🗿

[neat apex]

Yeah, an test jugded by like Sonnet will be better than jugded by humans, mostly because it is smarter than a average human istead consistency

[ocean vortex]

This is the most ridiculous thing I have read this week, no offense

This is like asking some entity to launch investigation into itself

LLMs are terrible at self evaluation

[neat apex]

Anyone is bad at self evaluation, if you swear something is good, you will judge it being good and do in that way

[ocean vortex]

The problem here is either you don't understand AI at all, or you are having issues trying to communicate it.

[neat apex]

The issue is not even AI, its about self judge

[ocean vortex]

yeah he can't even grasp that, true lol

[wispy quarry]

Can I only generate videos here in the server, or can it also be done on the website?

[ocean vortex]

That's too much schizo for me to engage.... all this 'psych' world 🗿

Classic example of someone trying to apply something he studied to different incompatible fields. You are not ML engineer my man...

[neat apex]

An "Self Judge Benchmark" that compares the level an AI check it self perfomance comparated to an actual realiable test would be an a interessing number

I trust Claude can jugde it self but not Gpt for example

[ocean vortex]

How would you even make the AI check itself without making ir reference some external human data? That approach is flawed by it's very design

And you can't make it generate test questions either cause it's never gonna come up with something it can't answer correctly. It needs to know the answer for the question to be valid

Generally from what I saw, even making it come up with the hardest questions it can answer is problematic. It's gonna sway into making the questions it can answer very easily. Because it is creating them based on what it knows and understands. To push it out of it's comfort zone you need human written/curated problems or tasks

[fiery gull]

Yep, just here 👉 #1397655624103493813

[prime mulch]

Does anyone love gaming and read manhwa especially cultivation manhwa

[gaunt spade]

idk what manhwa is

[prime mulch]

Korean comic

[gaunt spade]

oh damn

[prime mulch]

If you watch solo leveling that's based on manhwa

[gaunt spade]

tbh i dont watch anime

or korean comics

[prime mulch]

There is a game released yesterday

Thats a f2P game

If you love manhwa you should play that game

[keen beacon]

yo guys

what is familiar to chatgpt 4o in websites ai

like really really familiar

good for coding and stuff

[stray aspen]

gpt 5.1 high is a freak

i think its as good as gemini 3

[keen beacon]

anybody!/

[zealous sparrow]

this is python?

[stray aspen]

no its html

[zealous sparrow]

ah

[cloud zinc]

rip gemini 3

[neat apex]

Nahh, Gemini 3 will be better than that

I see Gpt 5.1 high is just like Gpt 5 ultra high since they adjusted the reasoning effort

[hollow imp]

Ok Claude models are literally unusable without Claude max

16k and 32k reasoning tokens are not enough

Gpt 5.1 extended thinking and gpt 5.1 high are way way ahead

[gaunt spade]

no its NOT

not even close

have u tested the new Gemini 3 checkpoints on AI Studio, within the A/B testing feature

[fleet lintel]

yeah, it's not even remotely closed.
not saying gpt 5.1 is bad but Gemini 3 is just another league

[gaunt spade]

gpt 5.1 can compare to gemini 3 flash

lol

maybe flash lite

[fleet lintel]

i wont be surprised that 3 flash == 5.1

[gaunt spade]

i have bigger hope in 3 flash than gpt 5.1

[wheat onyx]

i guess we'll get a much better idea of capabilities this week. im curious

[gaunt spade]

LOL

[fleet lintel]

I am just concerned that Google will raise price of Gemini 3 because it is soo much better

[stray aspen]

gemini 3 pro is nuts

[fleet lintel]

I think only pro is launching this week

[gaunt spade]

its voodoo bro

[fleet lintel]

i think gemini 3 might demand a premium.. but we will see. if google wants to earn extra $ or want to be super competitive

[gaunt spade]

and gpt 5.1 will allow u only 5 prompts on free plan, while you'll get much higher limits with Gemini 3 Pro

lol

[fleet lintel]

that was always the case

[gaunt spade]

they can, unlike openai

LOL

[hushed terrace]

Hi, please remember that english is the only language allowed in chat channels 🙂

[bleak lake]

Guys Sherlock alpha which ai is this?

[stray aspen]

nothing important

[gaunt spade]

they already did with 2.5 pro for 6 months

[bleak lake]

on open router, someone said it's grok 4.2

[gaunt spade]

thats xAI model

sherlock alpha

[gaunt spade]

its from xAI (Grok)

[stray aspen]

yeah it sucks anyways

[hollow imp]

Grok is the worst

[gaunt spade]

idk why grok sucks so bad at coding

lol

[bleak lake]

ah then grok 4.2 like they said

[fiery gull]

Nah bro, the grok 4 fast in the release I thought it really good in my use

[hollow imp]

Your standards are very low

[fleet lintel]

meta is the worst. Grok is fine but has some strenght at least

[gaunt spade]

even the worst chinese model beats it

[stray aspen]

gpt 5.1 is nuts

but i liked lithium flow more

[fiery gull]

I found the grok 4 fast better than the Chinese models at launch, but as it launched the glm 4.6 then it is useless now

[gaunt spade]

yeah glm 4.6 is a miracle at such low price

[stray aspen]

who uses grok bruh

[gaunt spade]

for its coding abilities

[stray aspen]

just shut it down already

[fiery gull]

Why some people said the gpt 5.1 dont is better that 5? I fell a huge improve

[hollow imp]

Models like that are out of conversation

[bleak lake]

grok is decent but nothing exceptional as far as I've tested. At least softer guardrails than open ai

[gaunt spade]

elon musk gonna hire some robots to your house

🤖

[hollow imp]

That's not decent nibba

[zealous sparrow]

you know what sucks with 5.1 tho

It has the same damn style

every time

[stray aspen]

yes

it lacks creativity i noticed its all the same thing

[gaunt spade]

no variety

[hollow imp]

@stray aspen ask your yupp ai to get gpt 5 pro api

[fleet lintel]

i have been wrong multiple times about Elon.. I wont make the same mistake by counting Grok out

[stray aspen]

unlike gemini 3

[hollow imp]

They are so rich

[stray aspen]

which gave me different style

[fiery gull]

About grok 5, do you think it will come really good?

[bleak lake]

i mean it's not completely trash

[hollow imp]

Atleast it doesn't hallucinate as much

[stray aspen]

and it likes putting a bunch of text so much

[gaunt spade]

also u can control the temperature with gemini models, which means different results each time

[fleet lintel]

is it 5.1 ? what is the prompt?

[bleak lake]

Atp grok 4.2 is coming out before gemini 3

😹

[stray aspen]

its gone lol

they removed it

this the prompt

[fiery gull]

Gemini 3 never will be released 😢

[fleet lintel]

2 more days for gemini 3. atleast we dont have to wait too much anymore

[bleak lake]

😭

[echo cosmos]

Grok 4.2 is worse than Gemini 2.5 Flash

[fleet lintel]

18 nov.. mark my words

[bleak lake]

2 more days 🥶

[fiery gull]

Just december

[fiery gull]

Lol 🤣

[ocean vessel]

Yesterday I tried deepsider.ai and for now it offers free Sora 2 (10s) ai video gen 👇

[stray aspen]

scam labs ai is better than grok 4.2 lol

[fiery gull]

Scam labs? Why? Because the "" mpu"" 🤣

[echo cosmos]

Deepseek v3 crushes Grok 4.2

[fiery gull]

We need more chinese spies

[stray aspen]

gpt-3.5 turbo crushes elon musks clanker

[hollow imp]

Yupp ai is removed?

[bleak lake]

Qwen is better than grok atp

[gaunt spade]

scamlabs is actually glm 4.6 with Claude writings

[hollow imp]

Yoo

Maybe every ai company uses chatgpt models to train their model but grok doesn't do so because of elon musk's grudge that's why they are so behind?

[fiery gull]

Serious? 🤔 Lol

[zealous sparrow]

isnt gpt 2 and gpt 3.5 dead

[gaunt spade]

well they actually said "we pulled some weights from qwen 3 and glm 4.6"

[fiery gull]

Gpt 2 is open

[zealous sparrow]

oh its opensrc

[fleet lintel]

what is scamlabs?

[gaunt spade]

movementlabs

[fiery gull]

Movimentlabs

[echo cosmos]

Even Qwen3 4B is better imo

[bleak lake]

😹

[gaunt spade]

grok providing the worst closed source models ever

[bleak lake]

Even qwen 3 4b Q2

[bleak lake]

real

[echo cosmos]

No joke, that's real

[bleak lake]

vibe thinker 1.5b > Grok atp

[fleet lintel]

grok is not that bad... its not the best but its ok 🙂

[zealous sparrow]

I feel like gpt models are better at python than html

[echo cosmos]

i didnt test it, is it good ?

[fiery gull]

Is soo massive the 4b 2507 with only 0.4% of qwen3 max even be a good model

[bleak lake]

for a small model its definitely worth it i'd say

[fiery gull]

I never tryed this new model, is multilingual?

[bleak lake]

good reasoning

[echo cosmos]

that's the goat of the slm by far for me

[fiery gull]

Too

[zealous sparrow]

If you are going to use a gpt model, i recommend for smething like pygame or python, because it aint lazy on that one [unlike html]

[bleak lake]

It does

[old igloo]

HEY GUYS

[bleak lake]

It even runs well on phone

[old igloo]

WHILE GENERATING A VIDEO CAN YOU NOT GENERATE A ANOTHER ONE?

CAN ANYONE PLEASE CLEAR ME ABOUT THIS IN?

[tulip tree]

https://tenor.com/view/ishowspeed-speed-reaction-try-not-to-laugh-bruh-speed-gif-15909423665204532371

[hollow imp]

@deep adder bro I asked the same gpt 5.1 model a very complex task in chatgpt ui and lmarena and in ui it completely declined my request and kept repeating system guidelines system guidelines but it perfectly worked on lmarena. So in api you not only have custom system prompt feature but also this annoying openai safety guidelines is gone?

[ocean vortex]

For chatgpt they are routing 'unsafe' requests to longer reasoning

[gleaming roost]

GPT 5.1 working for anyone?

[wheat onyx]

yeah?

[gleaming roost]

🤔

[tulip tree]

[gleaming roost]

For me, it just keeps generating the message and then gives an error.

[wheat onyx]

curious if OAI will release new models immediately after Gem3 comes out

[tulip tree]

I bet 10 dollars

[hollow imp]

There's nothing to do with safety for the prompt

Uhh discord is not letting me send it how do I send

[gaunt spade]

https://tenor.com/view/mods-low-tier-god-ltg-gif-4022085103641635431

[wheat onyx]

it works, but gave a dumb answer

[fleet lintel]

my guess is new improved image model..

[gleaming roost]

5.1 or 5.1-High?

[fleet lintel]

with some kind of meme-ability

[wheat onyx]

thinking. probably medium because im on plus

[hollow imp]

For us ppl using their models through api is the only choice

[hollow imp]

Then pick extended thinking and use nudge phrases

[old igloo]

HOW DO YOU KNOW IF YOUR VIDEO GENERATION IS DONE OR NOT?

[wheat onyx]

https://tenor.com/view/jake-gyllenhaal-angry-shake-mirror-scream-gif-16244849

[hushed terrace]

Hi! You´ll receive a message from the bot when your video is ready 🙂

[old igloo]

MAN ITS SAID 10MINS BUT ITS ALREADY AN HOUR STILL NOT GETTING

[tulip tree]

https://tenor.com/view/ishowspeed-speed-reaction-try-not-to-laugh-bruh-speed-gif-15909423665204532371

[wheat onyx]

I think if you say i will give your advice to a lawyer as the next step, it restricts legal advice a bit less

[hollow imp]

I don't believe you one bit

If you're so good then tell me how do I bypass openai's system prompt and guidelines

Very annoying

[gaunt spade]

GUYS NOOO

riftrunner got removed from lmarena 😭

[wheat onyx]

Kill the legal ban system prompts