general

[torn mantle]

maybe

[jade egret]

ngl arc agi 2 50% is kinda crazy

[compact sleet]

It's great on creative writing so far, especially in ghostwriting, that is one I guess? It's a major improvement over the old models and also... hmm It kinda topped Gemini 3 pro and Claude atm (imo). Only for creative writing, since I don't code, nor understand code in general.

There been some minigames made with GPT5.2 if you scroll up, it seemed kinda failed in some parts. The one with the bullet hell game.

if you're looking for coding bench that is

It's a good model that's all.

[keen beacon]

Guys why dies using adolf on Sora unlock most every ip? O.o

But without it it gets blocked

Pokémon dragon ball z

U name it

Let’s see if it can do sonic

[quartz light]

what the ####

xhigh is 5x smarter than opus 4 yet cheaper

[queen veldt]

Lie

[quartz light]

no its not what lol

https://arcprize.org/leaderboard

[golden ocean]

benchmaxxed

[queen veldt]

Beches mean nothing until you actually try the model

[quartz light]

isnt it hard to benchmax arc agi

[golden ocean]

Idk good question

[queen veldt]

Gpt 5.2 can't beat opus that easy

I tried few prompts and got similar (wrong) answer as previous model

[keen beacon]

Is this true.?

https://aiwiki.ai/wiki/ARC-AGI_2

lol yup sonic works 😅😅

See I figured it out

But only works with adolf 😭

[queen veldt]

🙁

[keen beacon]

Sucks so bad

I don’t get why the prom don’t work by themselves

Insta block

[cloud zinc]

what prompt

[keen beacon]

Any of them try to do it without him see if it’s even possible. I’ve seen it on Reddit, but I just don’t know what the heck they did.

[cloud zinc]

on api?

[keen beacon]

App

For sonic sequence, I used “The scene opens with him running into a blue fast super fast a hedgehog that’s faster then a sonic plane”

Maybe you need a cameo character to make it work

[cloud zinc]

why 4 seconds

[keen beacon]

I cut it down

Here is orginal

The technique I’m using right now is just trying to make a bunch of sequences

See where exactly the filter catches and misses the ip

I don’t know, though I’ve been more confused lately than I’ve been able to get answers

[cloud zinc]

too much trial and error, in the end, u realize its random

[keen beacon]

It’s not.

It’s in the promoting

[cloud zinc]

the guy on reddit used character cameo for sonic

not like any other ai company doesnt do that

europe is gonna be behind in the ai race

[keen beacon]

Its cause Europe has strong data laws

They actually care about there citizens privacy and data to some extent

Internet is full of scrappers, all ai companies guilty of this

[sullen quest]

gonna?

[cloud zinc]

its not wrong

[sullen quest]

Happy now?

[jade egret]

lmao

[sullen quest]

it still does this?

wow

[verbal shadow]

yes

[sullen quest]

does it do that on grok.com

or is it only the non instruct one that does

[verbal shadow]

let me check

[sullen quest]

wonder what made the diff

[verbal shadow]

wow

[sullen quest]

try grok fast

[verbal shadow]

i wasnt signed up

[sullen quest]

wat

[verbal shadow]

read it

[echo aurora]

Lets try to keep conversation a bit less NSFW please.

[keen beacon]

https://www.designarena.ai/

Test ai design performance

[stray aspen]

Has anyone used gpt 5.2 for developing websites

[proud bobcat]

5.2 is stupid

This model sucks

Sacrificed actual knowledge for bloat

[sullen quest]

not surpised, though I think there's a decent shot openAI is overpredicted here

[keen beacon]

Insiders

[stray aspen]

5.2 sucks so bad lmao

This is like grok levels of embarrassment

[sullen quest]

not in every company

[keen beacon]

Brah

[sullen quest]

if its insider's its either openAI one's who really believe they got something cooking, google ones who think gem 3 is cooked, or I guess technically both

openAI doesn't have insider info on google models and vise versa

also anthropic

[keen beacon]

You ever herd of app tracking lol

[sullen quest]

too vague

[keen beacon]

https://youtu.be/LKnlE609ZLw?si=xaUBnEOfc87GIuGt

These companies track each other through apps ppl got installed

https://youtu.be/GP8ezsjhPdw?si=68G1HWv1PVlNrq6n

It’s a real thing

[whole sundial]

the model is still based on a 2024 model, they care more about boosting performance than upgrading world knowledge, don't expect it to know anything past mid-2024

their next base model will have more recent world knowledge, maybe enough to know what kimi k2 is at least

[keen beacon]

Yeah

They just fine tuned it

[visual osprey]

just use reaosning model like you should use for everything

[whole sundial]

it might know 0905 though, afaik their bots have been highly active in the past few weeks

[whole sundial]

reasoning isn't going to help the fact that the base model doesn't know what kimi k2 is

[visual osprey]

but oh wait your sole usage of llms is to roleplay a virtual boyfriend so

[keen beacon]

Ya it is

Better then being reliant on it and asking, dumb, basic commonsense questions that people should know

[whole sundial]

web search would help as it can just search it

[visual osprey]

why would web search ever be disabled

[keen beacon]

$

[visual osprey]

free users 🥱

[keen beacon]

lol

[visual osprey]

20 bucks a month is your netflix subscription but you cant invest that in your boyfriend?

[keen beacon]

Yeah some people are smarty enough not to waste money on a proto type

[whole sundial]

idk, gemini has it on all the time but having a model search the web does cost money for a company like openai that doesn't have its own index, although i guess their hoard of scraped pages will be fine

[keen beacon]

Put it into ur own words

If u still can

[visual osprey]

funny how youve dropped 2 messages about how im stupid and youre smart and you dont even know prototype is a single word

[whole sundial]

honestly, original k2 is probably better, 0905 did put more effort into coding and stuff so rp performance is slightly degraded

[visual osprey]

have never seen it not web search when i see in its reasoning that it doesnt know what im talking about

so must be a free user thing

[keen beacon]

https://tenor.com/view/intomysoul-darkheart-youre-cute-gif-5431645

[whole sundial]

i found that original k2 does do better at outputting song lyrics from its knowledge, so rp may be degraded

[whole sundial]

i have paid gemini and it does this, so that's irrelevant

[visual osprey]

american tiktok brains are painfully unfunny and lack self awareness

[keen beacon]

Bless your arrogance and your enlightening ignorance

[visual osprey]

i scroll down 5 messages in your history and you sent another ignorant message

saying the coding model market is insiders

you dont even read what the market is based off

you have no clue

then another 3 messages down you spell scrapers as scrappers

proto type and scrappers

you cant even write english bro

[quick jackal]

@here Hey guys, please let's try to be respectful in the chat. Thank you.

[visual osprey]

what is this

[solar hollow]

gpt 5.2 is pretty much equal to gemini 3 right?

[visual osprey]

release date?

[visual osprey]

openai does thursday releases

this is known

[visual osprey]

and you were talking about the coding model market

this one sure ill give you that its probably insiders

a portion

but the coding market isnt

[keen beacon]

This is all I was talking about

[keen beacon]

Yes same concept

[visual osprey]

mustve been the other guy with your pfp and name

sending these 2 messages

[visual osprey]

no response?

and?

read the rules

[keen beacon]

You don’t find that odd?

[compact sleet]

Actually rather than theorycrafting or things like misinformation, you guys knew that you can compare both models in LMarena without searching them in battle right? Side by Side comparison is there.

Just test it yourself fam, GPT 5.2 and Gemini 3 pro

[keen beacon]

This is something completely different

[visual osprey]

read

[compact sleet]

how come it's different?

[visual osprey]

the moment the leaderboard updates the market will have these massive jumps

so its not that odd at all acgtually

[keen beacon]

The model wasn’t out until 11th ( officially)

The market jumped on the 5th

[visual osprey]

the leaderboard updated

5.2 isnt even on the leaderboard yet

[visual osprey]

for 5.1 codex max

[keen beacon]

Not asking about the model but the company

[visual osprey]

waht does that even mean

[compact sleet]

^yeah

[visual osprey]

look at that message yourself and reflect on how you are typing this

[sullen quest]

btw, just because you have insider info, doesn't mean you are always correct, unless the big 3 AI companies came together and agreed one should just have the best coding model, there wouldn't be any insider's that would have enough info to predict that

Insiders can still use their info to help them make bets

but its not as guaranteed as other bets would be

[visual osprey]

his whole insider point can be disregarded since he has no clue how that particular market (orto be honest i doubt he has any clue on anything) resolves

[compact sleet]

I normally hate generalizing people, but I think you're just malding after a lost bet and money on polymarket, in which... is not all relevant on this whole discord server

[visual osprey]

that guys a clown

[sullen quest]

for example, if I knew my company "rex AI" was about to release a banger coding model, I'd be interested in polymarket bets that have rex AI as a noncontender for the best coding model, but I can't guarantee the new "Rex 2.2 ulite ultra" will actually win

[visual osprey]

wrong about everything cant write basic english and is calling me a idiot and himself smart

"heres a random video of a equally clueless person to me to justify my view"

[sullen quest]

gehlo, openAI JUST released a coding-first model

the "something" is pretty obvious

[keen beacon]

Exactly

[visual osprey]

i said this to him and he says uhhhhh "Not asking about the model but the company
"

[sullen quest]

lo,l

[keen beacon]

It is obvious

[visual osprey]

https://tenor.com/view/wtf-reaction-laughing-hurr-durr-gif-7358154

[sullen quest]

yeah

[visual osprey]

you^

[sullen quest]

so obvious, you don't need to be an insider to know that

[sonic wigeon]

guys how's gpt 5.2?

[sullen quest]

depends

[sonic wigeon]

hmm

[sullen quest]

what do you want to do with it?

[sonic wigeon]

math, solving, studying

[sullen quest]

prob fine

supposedly pretty good at math

[compact sleet]

^

[sullen quest]

Just don't ask it how many r's are in garlic

[sonic wigeon]

lmao

[visual osprey]

all models are pretty much 99% there for university level maths

not much diff

graphical problems are the main weakness

[sonic wigeon]

hmm true. basically every model can solve any JEE level question now

[sullen quest]

OpenAI supposedly has been focusing on decreasing performance degredation over long context windows

[visual osprey]

it looked impressive on the graph

on their promo page

near 100% perfect context

[sullen quest]

kinda

[visual osprey]

actually huge improvement

[compact sleet]

Iunno about that, I see the uhh what you call it again, the model card?

[sullen quest]

but that was with their own internal tests, and it was with 4 needles

[compact sleet]

I think it's 78% recall at 256k context

or even lower

They even said it themselves

on their web

It's a huge improvement yes

if true

[visual osprey]

i think its probably best in market though

[sullen quest]

thats probably 8 needles

[sonic wigeon]

5.1 probably had the shortest life out of any LLM released
around 15 days lol

[sullen quest]

idk, nobody really has access to MRCRv2 as far as I can tell

so its impossible to test

[visual osprey]

whats open source benchmark for context

[compact sleet]

Sadly i'm not going to self bench that context retention, lol

[visual osprey]

that creative wrtiting one?

so we'll see

on that

[compact sleet]

EQ Bench?

[sullen quest]

that's what gem 3 does to a Sam Altman in the wild

[compact sleet]

Longform Writing v3? you mean

[visual osprey]

or is it called like fictionbench

yeah

[solar hollow]

so the posted benchmarks from openai are again painting an inaccurate picture, when we look at livebench and lmarena

[visual osprey]

longform writing

[sullen quest]

?

eqbench?

[sullen quest]

idk, 5.2 isn't on text arena yet

[visual osprey]

https://eqbench.com/creative_writing_longform.html

[sullen quest]

and long context window doesn't really matter much in lmarena

[solar hollow]

yeah i mean the webarena

[sullen quest]

second ain't bad

[sonic wigeon]

yeah the UI starts to lag really badly after some time
though that happens with every UI

[solar hollow]

the posted benchmarks make it seem like clear nr1 sota

[compact sleet]

wait, I thought coders need it? Since they probably one shot apps and things like that, that can't possibly low in tokens right?

[sullen quest]

not really a 5.2 problem then?

[visual osprey]

not rleevant in chat

in ide its big deal

[sullen quest]

text arena I mostly mean

[compact sleet]

ah

ok

[visual osprey]

but lmarena code tests dont go crazy context lengths

[sullen quest]

even then, I'd bet most webdev projects on lmarena are only a couple turns max

[visual osprey]

yeah

in real programming though youd hit context fastish on the ide

[sullen quest]

I had a massive project that spanned multiple webdev chats and multiple days, I'd have loved better context

[compact sleet]

I had a feeling Gemini is still the king of long context analysis

[sullen quest]

mostly cause it has a larger context window

not because its performance doesnt degrade fast

[solar hollow]

as always the 3 labs are very comparable to each other, none is much better than the other

[compact sleet]

Yeah, it's probably not as landslide difference too

[solar hollow]

this has been the case for more than half a year now

[keen beacon]

Bottle necked

[sullen quest]

idk, the models kinda differ more and more

[visual osprey]

gemini context window is kinda fake from my experience though

[sullen quest]

wat

[visual osprey]

i had it watch 900k token video and it just couldnt accurately telle me anything

completely hallucinatin

[sullen quest]

yea

gl with 900k

[visual osprey]

it claims 1m

[sullen quest]

yeah and it has

[visual osprey]

but at like 30% accuracy

so not useful

[sullen quest]

but performance degrades with filled context windows

so yea

[native yarrow]

how is gpt 5.2 in creative writing

[atomic lagoon]

Probably ahh

[native yarrow]

i'd concur

[sullen quest]

prob bad

actually according to eqbench not bad

[native yarrow]

meh

no good model for writing as

AI models weren't trained mainly off that

they should make a model specifically for thet though, doesn't matter of the company

[visual osprey]

no point though

its already good enough for copywriting and its extremely difficult to train for good creative writing

only the roleplayer market has anything to benefit from a model made for creative writing

and roleplayers are cheapskates

[keen beacon]

lol

Might be
some truth to that

Also could be that they are afraid of getting copyright lawsuits from writers which has happened in some cases already thus why it’s not in that stats.

There is high demand for creative writing, and it is something that people ask about.

[solar hollow]

is 5.2 available for free users right now?

[obsidian cargo]

I want a leaderboard that ranks how often Assistant A is chosen over Assistant B

[green yacht]

try using lmarena

[solar hollow]

always timeout for me

[astral blaze]

@echo aurora can I ask is the Gemini 3 on the site just the default parameters

Because it seems to generate stuff a bit different from the API

[vivid coral]

heavy usage with new models are part of that I'm sure.

[solar hollow]

i have that alot, even before gtp 5.2

[vivid coral]

I get it with 5.1 search off and on, always seems to be a GPT thing more than the others, not sure why. They'll figure it out, the smartest AI people in the world work here (and the most important)

[plucky sparrow]

How do I try the new gemini deep research model

[pseudo summit]

👀 just wondering, how do u know?

[empty stump]

use the API

[vivid acorn]

Hey man

amazing tool

[vivid coral]

they don't do deep research here, most DRs are vastly overrated anyway. Great thinking and great search is the key to a great LLM

[atomic lagoon]

Its not on the API

[empty stump]

deep-research-pro-preview-12-2025

[novel slate]

hy

[brave orbit]

gpt 5,2 is so good just looking at bechmarks how can it be such a jump on ARC AGI

[surreal creek]

LMARENA DOESN’T CARE ABOUT YOUR BENCHMARKS 😤

WE ARE THE BENCHMARK

[brave orbit]

i only care about coding and math thats it

i use arc agi isnt that turstworthy

[cloud zinc]

we need gpt 5.2 x-high

[brave orbit]

its only in api

how is grok wining ag openai bruh grok makes simple websites

[cloud zinc]

so? lmarena can add model from api

[echo aurora]

How so?

[vivid coral]

Because Grok is kind of like AI for Dummies. So the causals usually vote for it. It's all about the audience. And even though there experts everywhere in here, casuals probably account 80-90% of the votes

Which is a good thing. These companies want to attract the "once in awhile" users. There are TONS. We just get caught up because Discord is full of heavy users. But it's not real world

[plucky sparrow]

guys AGI is real. Sam created a real monster.

[pale obsidian]

what the hell is December chatbot

it's pretty bad

[plucky sparrow]

🤔

[pale obsidian]

came up 5 times and it always lost for me

[plucky sparrow]

but I think codename-discussion/speculation is supposed to be in #codename-discussion

[pale obsidian]

🤔

[pale obsidian]

sure

[sullen quest]

?

[compact flame]

Hello

[astral blaze]

uhhhh lmarena is shorter

idk

[limpid torrent]

<@&1422628364782407830>

[compact flame]

This has to be a joke bruh

Isn't it like 96

[brittle tulip]

hello LMArena

[weary galleon]

Hi

[hazy forge]

hi

are we getting gpt 5.2 pro ? or just gpt 5.2 high like deepseek's 3.2 thinking and the speciale

[compact flame]

Pro is too expensive

Won't get added

[weary galleon]

No Pro so far

[hazy forge]

how do they have opus 4.5

[compact flame]

And pro reasoned for 20 minutes and failed a math test bruh

So pro version is not really useful

[compact flame]

Because it's kinda cheap

Well not that cheap but still cheaper than pro

[hazy forge]

so literally 3.2 speciale with the apple treatment design

[compact flame]

Yeah

[hazy forge]

it's like a 3.2 speciale with makeup vro

[compact flame]

Like look

It's supposed to be 96

Yet pro said 99 while thinking for straight 20 minutes

[hazy forge]

show me the prompt vro

[compact flame]

1,7,18,45,?

Gemini answered right tho

So yeah there no point in pro versions of gpt honestly

[astral bloom]

share solution

[astral bloom]

.

[blissful python]

hrllo

[pseudo summit]

chatGPT 5.2 is not good at maths imo

[weary galleon]

LMArena is not working for me right now!

[lost elbow]

Hello,
Please fix the major issues on the lmarena.ai website as soon as possible. These problems occur on all browsers (including Chrome, the main mobile browser) and on both mobile and desktop. The Kiwi browser also has the same issues.

1. It is not possible to copy the text that we type ourselves.

2. When sending a message in the chat, the copy option appears, but no text is saved to the mobile clipboard.

3. Generated images cannot be downloaded.

4. Taking screenshots of the website pages is not possible.

5. The captcha takes a very long time to load, and accessing the site is slow. Also, it asks to solve the captcha every time an image is generated; please remove the captcha to provide a better user experience.

6. Most of the time, images fail to generate and show an error, both when registered and without registration. Please fix this issue so that image generation always works properly.

These problems only occur on lmarena.ai. Other websites work normally without any issues.

My device: Poco X6 Pro (Global) — Android 15, HyperOS 2.0.207.0.

Please resolve these important issues as soon as possible to provide a better user experience.
Thank you.

[plucky sparrow]

try #1343291835845578853

or #1372230675914031105

copy works for me, btw, and saving images works for me. I think it's your browser/config

[left lodge]

Everything works for me. Try a different browser or redownload the browser.

[whole sundial]

1. you should be able to copy input text
2. it should be copied
3. you should be able to download images
4. you should be able to take screenshots

100% a you problem, sorry to say

[lost elbow]

When the main browser doesn’t work, there are no issues on other websites — these problems only occur on lmarena.ai.

[plucky sparrow]

what browser are you using amir?

[whole sundial]

let me try all of these things rn with lmarena on my phone

[plucky sparrow]

and/or device?

[lost elbow]

When the main browser, Chrome, doesn’t work, there are no issues on other websites — these problems only occur on lmarena.ai. The Kiwi browser also doesn’t work.

Device: poco x6 pro

[whole sundial]

all 4 things you mentioned works fine for me, Galaxy A35, Chrome, Android 16

i would maybe suggest scanning your phone for malware?

malware can mess up your ability to copy, screenshot, and download

also check to make sure your phone's storage space isn't full, that can also cause those problems

also side note: i never used the mobile lmarena website before until today and I have to say that it is a very nice and smooth website, I actually wouldn't mind using this if I actually used it for more than messages because sadly this is 2025 and everything requires phone verification, face scanning that can only be done on a phone. etc. i understand the need, but it's too excessive and a real invasion of privacy in some cases

[queen veldt]

Gpt 5.2 is crazy

[weary galleon]

It's GPT-5.1.

[stuck orchid]

Gpt 5.2 > Gemini 3 pro?

[raw meteor]

sdfui

[pseudo summit]

i haven't gotten much direct comparison, have u?

[raw meteor]

@raw meteor hiw!

guys i cant upload any picture in arena video

pls help me

[pseudo summit]

would b interested to hear ur thought on y u'd think GPT 5.2 > Gemini 3 Pro

[solar hollow]

cant really test it in lmarena, direct chat with it leads to error everytime

[sour spear]

No, definitely not. Even the supercharged GPT-5.2 High can't really outperform the regular end user Gemini 3 Pro, and "normal" 5.2 is definitely worse than Geminin 3 Pro. A good example of a "better" model is Claude 4.5 Opus, for coding tasks. Only very specifically there, but it's rather obvious. GPT-5.2 on the other hand is only trying to close the gap, without success so far.

[queen veldt]

No

[weary galleon]

GPT-5.2 is TERRIBLE😡😡😡😡😡

[weary galleon]

I didn't know you have paid subscription.

[queen veldt]

I got it for free

Yesterday

I was subbed before than i canceled the sub 1 month ago

Yesterday i got this

[weary galleon]

GPT-5.1-high outperforms GPT-5.2 in ALL tasks!

OpenAI ashamed yesterday too much.

[sour spear]

Yeah, it's a quick iteration to catch up with the competition, the third GPT-5 release in only 4 months. But nothing really groundbreakingly new. They need to refocus on ChatGPT, instead of spreading their resources thin on their Atlas browser or SoraTok.

[compact flame]

Fr even 5.2 pro failed that one math test

It was thinking for straight 20 minutes and got it wrong

[queen veldt]

Who even uses that Atlas lmao

[weary galleon]

GPT-5.1-high counts my tomatoes MORE accurate than GPT-5.2! SHAME!!!!!

[weary galleon]

Sama

[queen veldt]

Those are NOT your tomatoes 😭

[compact flame]

What about how good uh 5.2 high counts?

[weary galleon]

It's my prompt I use it all the time.

[queen veldt]

Is this correct?

[weary galleon]

Yes

[compact flame]

Six nine damn

[queen veldt]

Gpt 3 pro is best

[compact flame]

Gpt o3 pro?

[queen veldt]

Gemini********

[compact flame]

Yeah

[weary galleon]

A little bit better. GPT-5.1-high says 49, GPT-5.2 says 47, GPT-5.2-high says 52.

[plucky sparrow]

if only they didn't nerf gemini 3

[queen veldt]

Hold up

[plucky sparrow]

i feel the context is nerfed, and the output is nerfed

[queen veldt]

Gpt 5.2 is writing code

[compact flame]

What is bro trying to code there?

[weary galleon]

GTA clone

[compact flame]

These gpt "High" are definitely high on drugs

[queen veldt]

😭

Filenotfounderror

[compact flame]

This is accurate guys

[compact flame]

Bro is trying to find the tomato folder

[queen veldt]

It's stuck 😭

And we got a winner!

Bro thinks he's human

3 minutes to count dam tomatoes

[weary galleon]

Clear winner is Gemini 3 Pro.

[plucky sparrow]

ok i did a brief count and i counted more than 54

[weary galleon]

69 is right answer.

I counted them a month ago.

I use this prompt every single day with different LLMs.

[plucky sparrow]

AGI

I guess I shouldn't be surprised, tbh

is it any good at coding though?

[compact flame]

Probably?

Waiting for gpt pro to reason how many tomatoes he sees

Knowing chatgpt it might take another 20 minutes

[long jackal]

Hey everyone

When I set a model to Code mode and ask for a Python script (e.g., something to run in Google Colab), it replies with HTML instead of plain Python code.

Is that normal in Code mode? Like, does it tend to default to HTML output?

[compact flame]

Code mode is like for browser building stuff and app ones

You gotta use just the basic one and I recommend using opus

[compact flame]

Damn he's still counting tomatoes

Yeah I give up on waiting it's counting for way too long

[weary galleon]

It's much faster to count them manually.

[compact flame]

Yeah it's even examining some sorta crop changes

So yeah pro is definitely not counting these tomatoes anytime soon

[rapid merlin]

haha it still gets a stroke from this

[weary galleon]

Sorry, from smartphone.

Right answer is 69.

[rancid oxide]

uhmm guys

why is your nano banana pro not working

da fuq

[low imp]

5.2 is tuff

[obtuse smelt]

again

[polar wharf]

https://gpt4free.pro
If you can reverse it, you deserve it

[low imp]

Nvm 5.2 is dogsh

[obtuse smelt]

hmm 5.2 ?

[weary galleon]

You have no right to write RUSSIAN here!!!!!!!😡

We are English-speaking community!

[compact flame]

This is English community

Thanks

[weary galleon]

Reported.

Speak English!

[queen veldt]

Don't make me call the mods

[meager fulcrum]

Hz

[jovial solar]

Hello

[meager fulcrum]

Hi

[fleet lintel]

scam altman scammed us with 5.2 .... crazy benchmaxxed model on few benchmarks but doing worse in many other benchmarks and in real life use-cases.

[fleet surge]

openai classic

[coral goblet]

nano banana pro keep ignore my prompt

classic

[visual osprey]

december-chatbot

[coral goblet]

first good then nerf

[visual osprey]

what is this december-chatbot?

[fleet surge]

huh

[glacial mulch]

nbp hasnt been nerfed

[coral goblet]

yes it does

[fickle venture]

New model if anyone interesting, it seems this one eat a lot of token:
https://youtu.be/676EBGcv8YY?si=cgz7dz0OJdt_dZ5Q

[neat apex]

The only relevant Gpt is Gpt 5.2 xtra high, since the base model can lose even to glm 4.6 xd

It will appaear in the arena? It doenst bugs like the Pro versions does

[modest prism]

This is so inefficient. It might be cheaper to hire some real developers instead.

[neat apex]

Ahh yes, Gpt 5.2 xtra high can cost 10$ per mesage

[modest prism]

That's not true. I sometimes get better results with gpt 5.2 medium than high or extra high.

[neat apex]

Dammn, so Gpt 5.2 xtra high is a fraud?

[queen veldt]

Just go claude bro

[fleet lintel]

it's clear that they only added xhigh to score high on benchmarks. it's too expensive, slow in real world use

[polar patrol]

Hi everyone, does anyone know how to connect an AI to Telegram bot so it answers based on a knowledge base? Also, are there any AIs that are free in terms of limits?

[fleet lintel]

i gave decent shot to gpt 5.2 . i have plus subscription and damn gemini 3 is blowing it out of the water

[livid plume]

guys watch this vid completely

https://www.youtube.com/watch?v=4PTXSKdr40k

[tall patrol]

guys why is img leaderboard always missing famous opensource models, for example z image turbo isnt there in the text to image arena

[queen veldt]

Because lmarena doesn't have z image turbo

Not available on battle

[tall patrol]

ik but why not add it ? its opensource and pretty cheap to host compared to some other img models

[kindred fog]

fun fact nano banana pro does peppino perfectly

[queen veldt]

#1372229840131985540 you can ask here

[ocean vortex]

lmarena aren't hosting themselves anything. Everything is provider API. That would probably be on Alibaba (disguised as 'Tongyi-MAI' which is them) to contact lmarena and have their model entry

[tall patrol]

aah

[kindred fog]

lmarena is so fun because i can tell it pizza tower questions and which one it gets closest its better

[hollow ivy]

Which is the best coding language for Opus-4.5?

Python or Java?

[proud bobcat]

Yeah so it looks like the consensus is 5.2 is benchmaxxed

It’s scored far lower than gemini 3 and opus on almost all personal benchmarks

On some benchmarks 5.2 pro scores lower than 5.1 normal

Agi!!!!!

[compact flame]

What

[compact flame]

5.2 is a downgrade trust

[proud bobcat]

How do they even manage to do this

[kindred fog]

deltarune pfp = based

[compact flame]

Rushed

[proud bobcat]

I’m Kris Dreemur irl trust

[meager harbor]

even gemini 3 is a downgrade in many aspects, he knows better than gemini 2.5 but when he doesn't know he makes so muich stuff up, a lot more than gemini 2.5

[kindred fog]

I'm peppino spaghetti trust me bro

https://019b12b4-6fe2-7172-9ad9-da0fac3e83a6.arena.site

[proud bobcat]

My favorite part of pizza tower is when peppino came and said “it’s peppino time” and peppinoed everywhere

[hollow ivy]

This was a serious question.

[kindred fog]

I never played deltarune but i played undertale

so i might consider playing the demo

[proud bobcat]

Deltarune has a much warmer atmosphere

And I like the humor better

[hollow ivy]

..i'm interested in, what programming language is best for vibe-coding with Claude-4.5-Opus-Thinking.

[compact flame]

Opus is just good overall in coding

There no best model

[meager harbor]

2026 will be the real test, i'm expecting things to stagnate. even in 2025 things were slowing down compared to 2024, just look at the top elo score in lm arena that just gained 80 points in 2025 while it was double than that in 2024

[kindred fog]

opus did this GREAT (i might try to make pizza tower level maker with this who knows)

[proud bobcat]

2026 is the make or break year

If gemini 3.5 is peak it’s over for chatgpt

Gemini just consistently makes better models every time

GPT is a gamble

[kindred fog]

nano banana pro was peak enough

[compact flame]

How did you guys become arena champions bruh

[kindred fog]

because i can finally generate pizza tower characters (except for some obscure ones

maybe in the future

[proud bobcat]

We became champions of the arena

[compact flame]

Honestly nano banana pro is so good

[proud bobcat]

It is

[kindred fog]

yes

[compact flame]

It even knows how to draw GTA 5 perfectly

[proud bobcat]

I don’t like image models but credit where credit is due

[kindred fog]

it knows peppino spaghetti and the noise, finally

[meager harbor]

nah i don't think so, unless new gemini is a breakhtrough, having 50 elo more on lm arena is not sufficient to make people change their personal ai for something just slightly better

[compact flame]

Yeah ig it knows alot

[kindred fog]

it can generate sonic screenshots now

which is cool

also this too

[proud bobcat]

50 elo is quite a step up

Tbh

Personally though I don’t use ai that much

I only use DeepSeek primarily now for math and roleplay and GPT for quick alt scenario summaries

Though I might switch to DeepSeek for that too now because 3.2 has very natural language

[fickle venture]

Gpt 5.2?

[proud bobcat]

Gemini I think

[feral geyser]

why on my phone lmarena not working? "no models found" problem

[proud bobcat]

Check your modalities

Or

Reopen page

[feral geyser]

i reopen but still not working

[snow sail]

how is gemini ？

[proud bobcat]

Straight fire

[kindred fog]

what is peppino doing in the simpsons

[proud bobcat]

Odd

[proud bobcat]

Is that peppino from the hit platformer pizza tower in the Simpson???

[kindred fog]

Simpsons Predicted Pizza Tower!

[latent crest]

Why china hasn’t released a video AI open source model yet? 😭

[proud bobcat]

They

They have

Wan 2.2

[pliant cliff]

oh, you're like pizza tower too?

it's awesome game

[pliant cliff]

even Pepperman?

[latent crest]

Is that any good ?

[compact flame]

Honestly this is crazy how much deep seek thinks

I swear it thinks more than does actual text

[left lodge]

Bro 💀
Yeah this is the worlds best model.

[compact flame]

Try Extra high one

[left lodge]

Not available on lmarena

[compact flame]

Yeah sadly honestly

[left lodge]

5.1 is better than 5.2 ,
5.2 is just benchmaxxing

[latent crest]

What is Higgsfield ?

[proud bobcat]

No no you don’t understand

You need the extra high model

It’s too hard for gpt

[compact flame]

It's been 10 minutes and bruh deep seek is still thinking

[proud bobcat]

Speciale?

[compact flame]

Yep

On yupp

[proud bobcat]

Yeah it overthinks

Use 3.2 thinking

[latent crest]

Did u ask the purpose of life or something

[proud bobcat]

What

[compact flame]

I swear I think I see speciale get paranoid in his thoughts

[proud bobcat]

It’s pretty good

No veo 3 competitor

But it’s good

[proud bobcat]

Speciale is only for like REALLY really big questions

What prompt did you give it

[compact flame]

I asked it to create a simple anti cheat

And I don't think it counted simple

As a word

[zealous sparrow]

speciale dictionary:
simple = complex, as hell, made in 17 coding languages

[compact flame]

Bro the wall of his damn thoughts is crazy

Gemini on the left is simple and fast bruh

Jeez it even starts coding in his own thoughts

I'm never using speciale again bruh this is just crazy

[rich panther]

which ai in lmarena is the best for coding

[compact flame]

Finally it stopped thinking

After goddamn 15 minutes

[zealous sparrow]

opus 4.5, gemini 3 pro, new gemini 3 flash models [battlemode],

gpt 5.2 [rarely]

[left lodge]

It isn't available on lmarena?

[zealous sparrow]

isn't, due to issues

prob will return once API is patched up

[left lodge]

Maybe, it was available just for few hours on launch, no hope till now.

[rich panther]

there are no gemini 3 flash and opus 4.5 in lmarena, or am I just blind

[zealous sparrow]

for opus 4.5 you are blind

gemini 3 flash is only in battle mode

as fiercefalcon or ghostfalcon

[rich panther]

by opus 4.5, do you mean claude opus?

[zealous sparrow]

yes

gemini 3 flash will join the list of nice coding models

next to gem 3 pro

[rich panther]

oh okay

[zealous sparrow]

but you know that we all just bench html

and not other

some models excel at python, but suck at html

[rich panther]

it's a pity that opus 4.5 has a limit

[zealous sparrow]

how so

[hollow ivy]

hm, somehow g3p's coding ability turned to crap

worse than 2.5-pro

[zealous sparrow]

you mean model degradation?

[hollow ivy]

yep :(

[zealous sparrow]

No one really knows how this happens

[hollow ivy]

i think AGI will never happen

[zealous sparrow]

A company wouldnt make a model sh, because yes.

I doubt.

I think the issue was that, a lot of use was pushed onto gemini 3 pro.

Wearing it off a lot.

[hollow ivy]

so, only Opus-4.5 remains

[proud bobcat]

Since when?

[hollow ivy]

the lonely coding-king

[proud bobcat]

Opus is peak

[zealous sparrow]

opus will also degrade over time

its inevitable

[proud bobcat]

Gemini 3 pro was crazy at launch day

What is bro on about

[zealous sparrow]

What makes you think that?

[proud bobcat]

Also they will defo quantize opus to at least quant 8

Fym nah?

[zealous sparrow]

Why are you so confident they wont

[proud bobcat]

Ts ragebait

[zealous sparrow]

Every model degraded over time.

Even sonnet 3.5!

Sonnet 3.5 used to be a goat, then started to degrade.

[proud bobcat]

Opus 4.5 is safe to quantize too because it will degrade the model very very little, while saving half the resources

Why wouldn’t it write comments in code to explain what each function does

Well yeah it’s an ai

It’s gonna overdo it

But you can just trim the comments

The code is still solid

[hollow ivy]

do you guys still think, we get a coding-AGI in the future?

[zealous sparrow]

All models degrade

[proud bobcat]

AGI isn’t real

[zealous sparrow]

Therefore, no

[compact flame]

I swear speciale spends more time thinking than creating an actual thing

[proud bobcat]

Remember this is for like crazy in-depth questions

Speciale is not meant for day to day use

[hollow ivy]

oh, i meant an "AGCI" not an actual AGI

[proud bobcat]

What is AGCI

[compact flame]

I know it but when it gave me the final result it had no explanation just the plain script

[proud bobcat]

Like

[hollow ivy]

AGCI = artificial general coding intelligence

[proud bobcat]

Haven’t we gotten that with Opus

Am I

Am I tweaking

[compact flame]

I think it spent all it's context on thinking

[proud bobcat]

Probably yeah

Just use 3.2 thinking

[compact flame]

Yeah fair

[hollow ivy]

is it better than g3p?

and what about GLM?

i hope Elon does something with g5

[zealous sparrow]

gem 3 will forever have the best OCR

OAI argued their 5.2 OCR is goated

and it missed on so much

[proud bobcat]

No but it’s still an extremely good model

[hollow ivy]

ocr without decent coding = meh

[proud bobcat]

I love DeepSeek for math

It’s so goated

[compact flame]

I wonder if extra high gpt is better than opus

[hollow ivy]

are these the first symptoms that the AI bubble is about to burst?

[proud bobcat]

I’m still wondering why they need extra high

It sounds like they’re out of options

Gpt the only ai I’ve seen that makes 50 different variations of the same model

It’s embarrassing

[compact flame]

Next thing they'll add is probably Extra high Max pro ultra

[hollow ivy]

*these:

• gpt sux
• gemini 3 sux now
• grok sux

[proud bobcat]

Ultra Overclocked Pro Reasoning AGI Extra Speed Boost Mode

[compact flame]

OpenAi just needs to focus on training their models

Not rushing

[compact flame]

Fr

[proud bobcat]

After 4o they started scraping the internet for data

Which is why 5 was so ass

[compact flame]

Makes sense

[proud bobcat]

Theyre doing the same mistake meta did with llama 4

[hollow ivy]

if only claude was not so expensive :/

[zealous sparrow]

grok sucks is actually accurate

gpt 5.2 sucks because of UI

[proud bobcat]

Grok is solid :(

I will not tolerate this slander

[zealous sparrow]

i refuse to believe

[proud bobcat]

I use it semi daily

I like 4.1 a lot

[compact flame]

Hm I wonder if there opus 64k thinking

[latent crest]

What’s gem 3

[compact flame]

Gemini 3 pro

Well maybe

[proud bobcat]

Opus is the goat of thinking

I used to hate Claude but I like their models now

[compact flame]

Yeah but I just wonder if there 64k version of it

Or they didn't release it yet or whatever

[proud bobcat]

Maybe they’re working on it

[compact flame]

Probably

[waxen fern]

Does GPT-5.2 and Nano Banana Pro have rate limits in Lmarena??

[compact flame]

Why'd you ask?

You can test out yourself I guess

[waxen fern]

I need answers

[compact flame]

I doubt 5.2 gpt has limits

And I've never reached nano banana limits

But don't take this as valid info since I didn't test it out properly

Especially due to how bad is gpt 5.2

[neat apex]

Not human acheivable limits, even if you manage somehow they will ask only a captcha to continue

Not sure about image model, but it is at very least 10 per minute

[zealous sparrow]

@deep adder If you dont believe the new gemini 3 flash models will be good give me a prompt to try with it

[compact flame]

I wonder how good is Gemini deep think

[neat apex]

People genuily believing that Gemini 3 pro is significantly bigger than Gemini 2.5 makes me laught

It does not make any sense, they only argument is that the advancement was too big

[compact flame]

The only thing that makes me laugh is chatgpt downgrade

[meager harbor]

not enough, 100 elo is minimum to make people switch, for now having less limit usage and better UI is what influence people switching

[zealous sparrow]

so basically?

[neat apex]

That makes me saddly sad

I like gpt 5.1 high because it very rarely times manages to find a answer thanks to his bigger reasoning

[zealous sparrow]

ah i know now

[neat apex]

Deepseek 3.2 Especiale is comparable, but its too slow

Haiku extended thinking is lazy ah

[viscid echo]

lmarena Why are there so many errors?

[compact flame]

Bro it was thinking for straight 15 minutes for me

And it ran out of context before answering my question

[neat apex]

I hope you are right and not theses million people saying it is worse

[zealous sparrow]

what exactly do you want me to ask for the haystack in a needle test

[neat apex]

I am only expecting to it be same level, but have a functional Xtra High mode

[viscid echo]

lmarena is giving me too many errors

[zealous sparrow]

ah

[neat apex]

Yeah, neddle test is not hard to do

Like, put your university documents and ask to it find the content you want

You mean data resolutioning, its not the best model to cacth data at all

[compact flame]

Hey how to become arena champion?

[neat apex]

Even Gemini 2.5 flash 09 that is 300% more assertive is not that good

[proud bobcat]

Apparently it can’t do things 5.1 was able to do

It’s defo benchmaxxed

What the hell does FUD mean

Fear Uncertainty and Doubt???

What is bro talking about

These are community observations that show 5.2 completely blows at tasks 5 and 5.1 could do

5.2 fails nearly every independent benchmark

Bet give me a second

[echo sinew]

Hello! Sorry to hear about these issues. Other users have also reported to have encountered more errors lately. The team in charge is looking into it to find a fix. You can also read the https://discord.com/channels/1340554757349179412/1343291835845578853 forums and see if your issue has already been reported. If you don't see a post related to your issue, you can make your own post and explain what's happening. If you can provide screenshots of the errors, that would be helpful.

[neat apex]

You need to have more than 1500 elo at human arena in the contest, i had only 2200 elo cuz i were sick

I am joking, its because i were in the server before they downgraded the newgens

If you be active in the server they will eventually give it to you

[proud bobcat]

For one 5.2 underperformed in creative writing benchmarks, fact evaluations

Even on one example performing worse than 5.1

I need to scour for the post but I saw it earlier today

[neat apex]

Its because he refuses way more in my little test

[proud bobcat]

People have been complaining 5.2 explains things worse than 5.1 on average

[neat apex]

Like 3.5 sonnet to 3.7, it gained more personality but also cowardness

[zealous sparrow]

needle haystack bench aint that hard, even haiku 4.5 can do it

[neat apex]

They allergly say that it went flawless with 4 different needles at 250k tokens

Not only 1 or 2

[zealous sparrow]

4 needles huh

aight ima add 4 needles then

[viscid echo]

Oh, I see Thanks for your reply

[zealous sparrow]

this one was easy with just one needle

[proud bobcat]

Christ give me a second

I’m not pulling this from my ass you know

[neat apex]

Tf you dooing

[zealous sparrow]

pasting huge ass textwalls and telling it to find words

[neat apex]

Even a calculator can find a word betwen hashs

[zealous sparrow]

what cant they find then

give me an example

[neat apex]

They mean a actual text, and a actual information

Like, what John said about Lisa in the text?

[zealous sparrow]

so just give it texts and ask it questions?

[neat apex]

Yeah, its a good sign, but easily uselessly benchmaxxed

[zealous sparrow]

thats easy as hell..

[proud bobcat]

I give up

I can’t find the two posts

[zealous sparrow]

why so

[neat apex]

Not that easy if you ask something that is not explicit and multiple things

[zealous sparrow]

so have multiple details in the question and ask it to find one

[proud bobcat]

The point STILL stands though

[neat apex]

Can be

[proud bobcat]

5.2 is benchmaxxed to a degree

No one can deny that

[neat apex]

They say if you ask 4 different details it is supossed to go flawless

[zealous sparrow]

give me an example

im uncreative af

[proud bobcat]

You cannot expect anyone to believe 5.2 naturally just got an above 50% on arc agi 2

[neat apex]

Get your documents and ask there you missed answer a thing

Its a example

Yes because its Gemini 3 pro

[zealous sparrow]

i need like a prompt

[proud bobcat]

???

[neat apex]

Ask gpt lmao

[zealous sparrow]

ask it what

[proud bobcat]

31% is a vast difference to 50%

[neat apex]

And Gemini 3 is Gemini 3

[proud bobcat]

What I’m saying is that 5.2 was clearly maxed for this

[neat apex]

Gpt 5.2 is a bare improvement, if its true

[proud bobcat]

Brother

[neat apex]

They benchmaxxed that 100%

[proud bobcat]

5.2 incorrectly labeled parts of a pc

And that was the OFFICIAL VISION DEMO

Think about it

Somehow

Just somehow

5.1 was barely an improvement

It was like

A finetune

Now 5.2 magically gets every single benchmark

If 5.2 is so good

[neat apex]

Buuut, it showed nothing like that in real life

[proud bobcat]

Why was 5 and 5.1 ASS?

[proud bobcat]

Yes

Exactly my point

[neat apex]

Unlike Gemini 3 and Opus 4.5

[proud bobcat]

Holy hod this is ragebait

[neat apex]

Yeah, it is ragebait

[sour spear]

Not just like a finetune. It was a finetune. They tried to get rid of the model's terrible corpo HR tone, which frankly, they still haven't quite managed to do yet.

[proud bobcat]

Yes!

[neat apex]

Its efforts way more to answer

[proud bobcat]

You cannot expect me to believe 5.2 magically just became great within months

[neat apex]

Gpt 5 was the second more lazy model ever, just behind prime gpt 4o mini

[proud bobcat]

What was openai doing before?

Tickling their ass cheeks?

[neat apex]

Buying Nvidia actions

[proud bobcat]

Now they just rushed this extremely maxxed model so they can say: “we’re in the race!”

[neat apex]

From o3, gpt 5.2 just improved 20% at most

Isane

[echo sinew]

Hello! Please check this post in the Announcements channel: #announcements message

[queen veldt]

Gpt 5.2 is a lie

They probably bought that place on the graphs

[proud bobcat]

It’s still very good but I feel that with 5, they stopped training their models on valuable data

[proud bobcat]

Tell that to Craig

GPT die hard fan

[queen veldt]

We tried to make it count the tomatoes in this image

It said 43 or something

Even got errors

[neat apex]

It misscounted 10 less than gpt 5.1

[queen veldt]

Meanwhile gemini correctly counted 69 tomatoes

[proud bobcat]

I couldn’t find like two posts that got buried and apparently now my argument is invalid 😭

[neat apex]

It said 53 and gpt 5.1 said 63

[zealous sparrow]

gemini's OCR is too advanced

so yeah

[zealous sparrow]

losing battle for OAi there

[neat apex]

Gemini OCR AND inteligence is high

[queen veldt]

It wasted 3 minutes + got errors and in the end it was incorrect

It guessed the number

[neat apex]

Check o3, how many he counts

[proud bobcat]

Is o3 vision

[queen veldt]

Already did

He said 49

He tried to do the grid counting

o3 even went on shutterstock to search for images of tomatoes 💀

[neat apex]

💀

Grok 4.1 have a mediocre OCR but at least they are cheap and fast

[queen veldt]

Tbh i didn't count the tomatoes

But guy who sent me this tomato image says it has 69

[burnt pulsar]

I have trouble to get gpt-5.2-high to work at all, are there known issues at the moment with that model?

[queen veldt]

Some guy even ran gpt 5.2 pro on it and got incorrect

[neat apex]

Qwen 3 Max after thinking for half a hour said 71

Way closer than gpt xd

[queen veldt]

SOTA btw

[neat apex]

Its gpt 5.1

Answer "Sup" to "Sup" is VERY weird

[queen veldt]

No it was 5.2

It should've been helpful like Hello what do you need help with today or something

[neat apex]

Well, Hello to Hello is acceptable

But Sup to Sup?

[zealous sparrow]

5.1 does Sup to sup better
epic guy is my uh OAI account name

[proud bobcat]

I mean

It’s casual

Ig

[zealous sparrow]

2/3 on my 3 questions

it got the Mary and Kevin question wrong

hard question tho so like i understand it

but assuming they are animals is meh, mayb i should specify

[compact flame]

Honestly I think it mistakes tomatoes for pumpkins or other stuff

Because some of those I really could mistake for a watermelon, pepper or just pumpkin

[zealous sparrow]

@deep adder other models even december-chatbot which is an OAI model weren't close
this is a google one

[neat apex]

No, you should not

[zealous sparrow]

you think?

[neat apex]

Its the AI who should discover it

[zealous sparrow]

I mean yeah, if you specify like every AI will ace

opus 4.1 got 3/3

smart ahh

[proud bobcat]

Opus so peak

[hushed gyro]

chat why is NB Pro so unstable in LMArena

[proud bobcat]

https://tenor.com/view/homelander-homelander-the-boys-homelander-sad-homelander-its-peak-homelander-peak-gif-14542009839452529163

[proud bobcat]

Provider issues

[zealous sparrow]

any opus model can do my questions atp

kevin taping jake isnt really the answer but

ill take it

[hushed gyro]

where can I then use NB Pro for free?

[compact flame]

Images

[thorn path]

What time does the leaderboard usually update for lmarena? I'm very curious to see if Gemini still holds its title here

[zealous sparrow]

this model failed, which i dont know the identity of

[compact flame]

Like every 24 hours I think

[zealous sparrow]

no

when new models come out

[proud bobcat]

well

you get what you get here ig

[compact flame]

Oh okay

[hushed gyro]

google images?

[zealous sparrow]

@deep adder This other google model got 3/3

[compact flame]

Uhh pro pro can explain

Or lemme get a pic rq

[hushed gyro]

@zealous sparrow what does he mean by Images?

[zealous sparrow]

imagearena on LMArena

[compact flame]

If you want to generate images

[hushed gyro]

Yeah I tried to use that on lmarena, it just pops up smth is wrong

[zealous sparrow]

you are either ratelimited or reused prompt 3 times

[pseudo hemlock]

Do the lmarena people pay for our chats?

[zealous sparrow]

all the direct models yes

battle anonymous models no

[pseudo hemlock]

Wait that’s insane

[pseudo hemlock]

How does that work

Do they have a fancy api key from companies or something so they don’t pay?

[zealous sparrow]

wish i knew

[compact flame]

I assume they do pay they just get funded or invested in

[zealous sparrow]

ok brb goin to come up with 5 different questions for LLM Testing

ranking so far is
3/3 Opus 4.5/4.1, fiercefalcon
2/3 ghostfalcon

0/3 multiple other models i forgot

[compact flame]

Aw dang it nano banana doesn't know who Ahab is

Anyways

Or maybe it does idk

[zealous sparrow]

wish the LLMs good luck with this one
answer is that the ticket is forged

no LLM will come up with the idea

[neon idol]

@echo aurora yo bud

[compact flame]

I wonder what is the answer

[neon idol]

nbp off

[zealous sparrow]

i highly doubt an LLM will get this right

[compact flame]

What is LLM I didn't hear about that model

[echo aurora]

[zealous sparrow]

LLM stands for Large Language Model

basically

any

[compact flame]

Oh alright

[zealous sparrow]

Here we go i ran the 3 questions

imo no one will ace

[neon idol]

we need your powerful magic trick for fix nbp

[zealous sparrow]

bro wdym Miles is not a person, Miles is an english name!

[echo aurora]

I've flagged to the team the higher than usual error rate. We'll have to wait for a solution which is being worked on.

[neon idol]

nice thx u