#general
1 messages Β· Page 333 of 1
screenshotted that watermelon
yea
why pressure on him
exactly
I touch grass btw
idgaf bro
@light sleet
exactly
honestly idk why u wanna screenshot stuff
no point
Cuz ur feline
who are u gonna send this to
@light sleet
the mods
OMGGG
hello guys
banana
wsp
watermelon what do you want from me
π
A screenshot of someone saying "Screenshotted that watermelon"
im not leaving the server or uninstalling the extension until banana says so
i dont know what to do
so im doing nothing
what extension
@storm dust Threatening to report me to the mods because I called him out for pentesting unauthorized
im there with no purpose anymore
@light sleet
what the hell is going on in this chat π ..
@surreal zephyr ALTMAN MOGGED HIM
read the conversation
buncha kids arguing abt some lm arena extension
Hi lol (I'm 6'10, I like matcha, and I wanna end period cramps)
dude it is illegal
Cuz they be like:
in a little bit of course I hope im in dont call that song
take it up to pineapple then, why r yall arguing in the public chat
i didnt know that but it was obvious
im waiting the right time
everyone SHUT
an extension that needs your browsing history to fix arena (Arena Team disapproved them but they're now spreading hate lol)
i even dmed to modmail
yap yap when spud
Sonπ
Whats modmail gonna do bud
theyre gonna ban you for pentesting
spud thursday
spud today pretty please
spud tomorrow π‘
spud yesterday
today
5.4 cyber
Blud gpt 5.5 π€
spud is 6.0
spud 1 month ago
Spud potatoππ
mythos release
Isnt spud 5.5?
Bro 5.5 was released tomorrow
no its people unable to read that new base = 6.0
I see
hey banana
ampro is sam altman undercover btw (hes pretending to be a member)
big fan sam
i know you got pinged a lot
BM MASS IGNORING
HAPPENING: RIGHT NOW
im sorry you were involved
EVERYONE IGNORE BM
who is bm
we need to talk in dms
The stoopid
damn
Bum
Pentester unauthorized
tbh i do the same but atleast i know the risks
you just admitted to committing a crime on a chatting app
well theres no proof xD
maybe i was sarcastic
fair
yes
dont expos
π€
I am sneakers o goat
tralalero tralala = blasphemy sneakers shark
You're just saying words atpπ
- Palestine Provoker
Sam when u gonna reveal yourself man π
bro its true
this chat is the definition of in heaven looking at hell
tralalaelel tralalla not human π
Its general
yes, chill
spud now
actually im cancelling spud
Tralalero tralala says a blasphemy in the original video
lelele
Spud also did
I touched dried grass
Bombardillo too
67
When will Gemini 3.1 Pro return?
Bro?
I'm crine
who wanna unauthorised pentest with me
no
NOO
You forgot cupcake
why
π
Bro i just made something so good wait
Im getting brainrot ππ€ποΈπ₯ππ§π»βπ¦²ππ»
spudalero spudala
real spud
Do u like spud potato
i prefer sweet potatoes
Sam, make GPT Jarvis to beat mythos ππ»
Smev, do u eat the fish
I'm the fish eat
Do u like coal
Son π
Bradar coal yourself bradar
wuthering wave characters or genshin characters
π₯
I need to lock in π
Cog Dock ππ»π
Bro ππ
bradar
Bradar
Bradar vat is this, delet delet
'Bradar what is dis comen" type face
I think I'm straight guys
is he having on argasm
no way
guys maybe yall shouldn't
Who wanna play
me
if mod 9, 10, or 11 wakes up yall are dead π
peakoexp13019
Why
8, 5, 16, my bad
Might as well take the risk
We only get to live once
π€£
we live in 1984
Man whose grown up plays this π₯
I don't have Roblox rn, that pic was 15 showers ago
Bradar vat is this, delet roblos π₯
π₯
dang grok is eons behind sora
@daring rock ur a bot
companionπ
bro is gona get executed π
What is happening
Tung tung tung sahurry and leave broππ
cuteeee
You just leaked my weak source
huh
no joking in big arena π₯
I love dogs.
But i have a dog named cupcake
do u like bacterias
noice
Sam
He can talk
I ate my dog and adopted a cupcake
where r u guys from, I wanna touch grass there
No one falling for ts
Spud spud sahur
Bro
And i use ts
Thats not triple t
Can we add 67 to global slur list and blacklist it
I still have other brainrots lilbro π€
Is that a cat
Cmon man
no
Why
Is that ur kryptonite bro
Hmm
scrolling thru my gif folder knowing one missclick would get my ahh banned π
(joke mods dont ban me)
sora 2 eons away from seedance 2.0
seedance 2 aint even real
Once upon a time
Bro
bro
use Websim bot
theres no seedance website
because its the best bot
whats that
it has free seedance
where
and sora
give
dms
π
wait
give
π
π
Idk, ask me
hey amigo
How this happend
What
Whats the algorithm
is it ok if i ask you something?
Idk
Sterling?
Whatt
could you please check my bug reports people claim that i am "pentesting"
i need your opinion
i know im not
Wait
U dont even know what it means and u just go straight to "I not do it!!"
uolol
whats a pen
pentesting unauthorized = finding vulnerabilities in stuff without authorization
Basically
oh like hecker
Are you chinese
no
ohlol have you ever heard of 401 status code?
BM have you ever heard of pentesting unauthorized 101?
because im sure you did
now zippy that mouthy shutty
this message explained your personality to me
im blocking you
omg u blocked me im so sad
Im crying now
Look at what you did ohlol you offended me
seedance is not even 1/10 as good as sora
ampro i dont know what that so called "guy" told you dont trust him
ur kinda right tho
he's a ragebaiter
Okay not sure why people are fighting with each other, but going to timeout everyone involved
pineapple before you do
please check my bug reports they are important
lmaoooo
api π€
shutdown by end of week iirc
sora actually looked realistic unlike all other
Oh damnn
dead
What's the difference between GitHub coplot and Microsoft copilot
microsoft copilot is just chatgpt and is bad
github copilot lets u choose models like claude or somehting, its not a model itself
i think
oh ok
Hii
Yea
nice, I don't have tho
Oh okay
Hey. You would have a way to check the code mode. It's giving an error to many people
Hi, can someone please explain why I can't use Claude Opur 4-7 thinking in text arena when I can see it in leaderboard?
they're involving me too idk why
Its only in battle mode since its extremely expensive its 25$ per million output tokens- all other opus models are not available either in direct or side by side you have to be lucky and catch them in battle mode
why my arena.ai not working or answering any ai
Something went wrong while generating the response. Please try again.
Trace ID: bb2e676a-dfef
Cx and its not even close
Can you drop the Trace ID in the #1417174113092374689 channel?
Fr
in llm arena this is without thinking ?
Yeah
Never gives a crap to their users except for the data they earn money from when they vote models
It's in Battle mode
So the model without the βthinkingβ feature is more powerful than models that do think, like the Gemini 3.1 Pro
Beo bring it to direct
Yo users are dropping massively because of it
We may in the future
More users using claude opus 4.7 means more model data means more money for you
?
These are what the leaderboards show. Sometimes just because it's a thinking model doesn't mean it's always going to be preferred
Still better then Slopus 4.7
what is slopus?
Opus 4.7
Popus*
Is Slopus 4.7 on Web/App
Good One
ok
c@light sleetare you the pineapple
Will Opus 4.7 be added to direct and side-by-side?
hell nah
Is clearly a Banana
Never
wtf
Sad
Maybe tomorrow
seems like gpt gemini and claude are down atm
This is useless: https://discord.com/channels/1340554757349179412/1466486650170245435
Why's seedream getting cucked??
Also, if it's not permitted, why is it an option?πΆ
Ping me in the thread you've been asking this in #ask-here and I'll get to you soon.
If you can share the Trace IDs in #1417174113092374689 that'd be helpful.
opus 4.7 reaches number 1 on the list
can not even use opus 4.7 as opus models got nuked off the site
????
It's in Battle mode, which is the mode that contributes to the leaderboards.
Can you share this screenshot in #1417174113092374689 ? I'll be sure to get to you when I can.
Non thinking model can be better than thinking model ?
5.4 pro standard (half the price total) vs mythos
Sometimes they can "overthink" a response
5.5 will mog all
6 spud
No need of Slopus and Myth OS
Even better
Native voice and video is peak
watch
Next thursday *
Spud is confirmed 6
tomorrow π
Theres no 5.5 π π
In my basement
sammy altman pls images v2 tomorrow
Images v2 might be part of spud
I think images v2 tomorrow
Not separate model
Then we would be cooked
Spud 6.0 is
Omni, new base
Spud will Mog All
gork 4.3 in beta guy
@echo aurora new announcement update, it was updated 1 min right after ur announcement I forgot to share xd https://019dab7f-4233-7287-81f1-8160e93f9ab7.arena.site/
Haha love it!
How do you manually update in? Just go back to Code Arena and tell it to add it?
yea xd
Lol nice
this time for some reason i couldn't copy the text so I wrote it manually too
as soon as the notification popped up
Asked glm to create new article lol
I wonder if there is a way to update it automatically
@echo aurora Just a quick question.
Ofc there are words that are blocked from the input prompt (I mean the tos), but what if the model itself outputs those blocked words? Do I still see the output?
i could not able to see claude opus 4.6 and 4.6 thinking in my agent from last 2 weeks
did anyone else also facing the same problem
Ofc there are words that are blocked from the input prompt
My understanding is it'll take into account context as well, compared to specific words it doesn't like (small difference but wanted to point out incase helpful information).
what if the model itself outputs those blocked words? Do I still see the output?
I don't believe so as there is a check on the input and output. At least there is for Image & Video, I assume same happens for Text
Yea
Recently some models have been removed from Direct and Side by Side mode. This was done to help ensure reliability and availability of Arena in the long term. You can find more information in this announcement: #announcements message
gpt image v2 is out
im not joking
Could you please take a look into https://discord.com/channels/1340554757349179412/1495352743555108864? Thanks.
linkknkn
if this is real then maybe this is why chatgpt is having issues rn?
but people have been occasionally getting image 2 in chatgpt for over a week now
banana disapproved
By banana disapproved I meanf
He disapproved me
Images v2 is indeed rolling out
first time in my life π
yeah i'm still getting image 1.5 though
V2 is not out
will probably come out tomorrow or thursday
keep in mind openai releases their image models usually on tuesday
It's prob rolling out by usa
Yeah can do
Gemini 3.1 and Claude 4.6 Opus are gone from Direct.
I think images v2 is rolling out
Pineapple . I think each chat should have an token limiter like exact number to know how long we need to know that it gonna end.
Pineapple what's the first thing youre gonna generate when gpt image 2 is out
I'll generate Sam Altman holding a pineapple
Gemini 3.1 and Claude 4.6 Opus are gone from Direct.
how about you?
ye they removed it because it was to expensive
It's so unfortunate
makes sense
oh yea kimi 2.6 is here
Qwen 3.6 max is out too
Seems to match Opus 4.6 performance
finally
3.6 Max Benchmarks
Wait for 5.5
yo sammy alternativeman I know u rolling out images v2
soon ;)
spud is going to be really fast btw
like fast fast fast
and smart as hell
new architecture
I saw Opus in the announcement and thought it finally came backπ₯²π
hi guys
Hello 
1.5
How fast?
hi pineapple im currently trying to make my own ai model
Kimi 2.6 came out
my goal is to.. get into lmarena leaderboards
EVEN if its like top 200000
Not sure how good it is
pineapple can you give me a h200 so i can train my ai ok thx /satire
Currently is added to Text and Vision btw, going to put out an announcement soon 
If it gets it it's 2
That's amazing!!! Be sure to let us know when you're ready
I'll check it out, keep up the good workπ
old youtube?
definitely 1.5
but 1.5 mightve improved a bit since last i checked, bc it used to be much worse than nano2
was I not timed out
Does anyone know which AI this is? It created me a beautiful perfect website. I cant describe how good it is (Its just for learning for a friend its all in german)
stealth prolly.
wait can you send image of the website rq?
it might take like 3 years mostly because im a teen chud that only has like access to google colab and stuff and whatnot
but ill try my best
as i said its german
it mightve have been better to just finetune a model but f it we ball
I wanted the simple desgin
i keep getting Something went wrong with this response, please try again., no matter the model i try
and its just too perfect
its kimi 2.6
i've tried quite abit
no wayyy
kimi 2.6 came out like hour ago anyway lol
I think that will be a crazy ai model
holy inspiration
no wayy what??
Classic "make a website based off of the prompt that you sent above"
oh go for it lmao
good luck friend
ty ty
yup 1.5, doesnt get abrams x correctly
βοΈ
idk if you or anyone here speak german but I think you could translate it from google chrome but just for you to see the website how beautiful structured it is:
https://019dabc1-8f3c-7598-bbc0-a5ab9fd63acf.arena.site/
Ive already created lots of websites with claude and so on but something is diffrent on this one.. I cant explain it..
even if its just that simple.. something feels 1000 times better than on other ai websites
whats the best model except claude models for frontend design
Can you report the Trace ID in #1417174113092374689 ?
nw done
Anyone know if opus 4.7 is in search battle mode?
Or how it compares to opus 4.6 search
opus is removed due to high cost
gpt models always do this weird glassy design which i dont really like
Not from battle
yea
yea me too
im a patient guy tho so i just tweak him to what i want properly
100% takes abit tho
No, not in Search Arena yet π
Dang ok ty pineapple
pineapple can i borrow a h100 from u plss /joke
if you even have one that is
How come some people have agent mode and some dont? I looked at the blog and it seems pretty vague like some people have it and some dont im still kind of confused
is sonnet good, ive been using codex mainly cause i tried sonnet once, token limit reached π
it's good but limited
Yea its amazing! The design is good, benchmarks are amazing its better than opus 4.5 i believe and it has become one of my favorite models
Even tho i have always disliked sonnet models until now
It's an experiment: https://help.arena.ai/articles/1811908126-arena-experiments-agent-mode meaning it's going to be random if you get the experiment or not.
Oh okay well lucky that i got it lol Is there like a page dedicated to exactly how it works? Like what models, what exactly it does etc. Or since its in beta there isnt really an article like that?
whats agent mode, doesnt say anything in this post and im quite intrigued/interested
Since it's an experiment we're not providing a lot of information about it. But the idea is users should use it as a multi-modal chat.
Hey folks at the Large Model Arena, when are you planning to add a ranking for model vision capability tests?
Why did the Opus models disappear from the arena?
Okay, thank you^^
They are still in battle mode but to expensive to host in direct chat / side by side
You can see the rankings for Vision Arena here: https://arena.ai/leaderboard/vision
Ty
Oh, sorry, I didn't notice that
Can I borrow a ak47 from u?
also the worst checkpoint of images v2 is rolling out
Not as good as duct tape
Seedance 2 guys.
Seedance 2 fast
that looks good
Ty
Na
It's no problem! Let me know if you have any questions about our leaderboards.
opus 4.7 when
In Battle now 
idk, sterling
yea but it's releasing by region
@echo aurora U heard the man, its time to ask the team!
hello @echo aurora i wanna ask why opus is no longer available in lmarena?
If it has been released in some regions, please release it now via direct chat in Arena AIππΌ
Whatβs the consensus on kimi k2.6
Exactly.
oh it worked
nvm
@storm dust I'm sorry.
Yo night fury
blocked people
Why u so rare to achieve
i wont unblock you again
Whats yalls fav ai model atm?
We don't need to be discussing this in general, asking to move on please.
fr? thats crazy
I didnt get image 2 yet
lmao
which is why it should be added to ARENAAAAAA!!
@echo aurora I did the model request hahaahahh
fixed text
look how everyone's model requesting image 2 now
shouldve used google flow there wouldve been no gemini tag π
Everyone model request it
NOOOWWWW
shouldve used arena
π
Shouldve use Google AI Studio
Nah wouldve taken too long, dont really like image generation there, im fine with google flow since it gives unlimited nano banana 2 generations
Okok
Image generation is not free there
Gemini 3.1 Pro is
When Kimi k2.6 :)
@echo aurora Can you start slowly releasing models like opus 4.5 and stuff
LET IT OUT NOWπππΌ
First, image 2
To Direct and Side by Side?
#ukunderground #dancehallmusic #newmusic
Based Lychee coming with this Big gy4t song DA DON OF LONDON prod. Mastercard2k
Thanks for Kunal for pretending to be Bobz and Thankyou Richard for being my cameraman!
Go listen to it on Soundcloud - https://soundcloud.com/basedlychee/sets/da-don-of-london
Like and Sub! Make sure 2 comment!
The direct chat model would be: gpt-image-2-high-fidelity?
Ye for image
If we make this change we'll be sure to put out an announcement
New title @echo aurora
Ok thanks
qwen crushing coding
Is it good?
i wonder how good kimi will be
according to the leaderboards yes
i have yet to test it myself
Ah, I thought you already tested it thatβs why I asked
Iβm testing qwen 3.6 plus right now, but I wanna know which oneβs better the max version or the plus version
@echo aurora News Update https://019dac15-4da8-7b09-82bc-fccad242b04b.arena.site/
Pineapple if u have time can u come to Chatroom
And set ur name to Pineapple or anything
In the middle of a couple things but I will later today 
Kk
yo pineapple do u know which models agentic mode uses
Since it's in the experiment phase I'm not going to speak much to the details about it, but us receiving these questions are helpful for us to understand what questions users have when they use the mode.
Hi pineapple can I eat you

Without actual announcement it's on lvl "trust me bro"
bro you can literally see people with it
Ignore how the tail is on the back of the car.
Can't unsee it
Why do I get session not found everytime... bruh
sir do you own arena.ai :0
β hes the cm
who owns it then π‘
John L LMArena
they should add image support to claude opus 4.7 thinking
That's very likely a connection issue. I'd recommend trying: different browser, turn off/on VPN if you're using it, clear cookies/cache
jokes aside its made by these people i think
Wei-Lin Chiang, Anastasios N. Angelopoulos, Ion Stoica
Yeah they're our founders + CEO & CTO
oki β
get me in contact with dem pwease
π₯Ί
Canβt access the website without VPN lol "π¨π³"
About what?
@urban herald hi
I wanna make arena.ai better π³
yo leako wsp
Tell us how in #1372230675914031105
wsp twintwin twinsicle
It was working before tho, dunno why not anymore
eyy twinsicle π
I mean I want to be a dev for it ik how to code lots of coding languages π³
eyy
Does the cooldown keep increasing everytime you reach the limit?
Yesterday I reached the limit and got 15 minutes I think, I reached again today morning it gave me 25-30 mins... I reached limit once again and now it says come back in 50 minutes... Will this keep increasing over time???
How does this work?
There is a limit-?
Never reached it π
You have reached your rate limit for gpt-image-1.5-high-fidelity. Please try again in 46 minutes.
Reached the limit 3 times in 2 days
how many can you generate before the limit
Btw why would you chose gpt 1.5 instead of nano banana 2
Idk I made like a couple dozen? Maybe like 15-20 images...
idk I just choose whichever came up first
Lol
Wow thats more thani expected
v2.0 mogs nano2 though
Well v2 isnt on lmarena yet sadly
It does for generation but I'd bet nano2 is better at editing already made images
nah v2 is better at everything
Why isnt your nose massive? nano 2 is inaccurate...
I havent used it yet but it could be, I cant wait to use it though
And I dont even like generative stuff
LMAO
Wow qwen 3.6 plus made this
I wish I was still friends with that dude, thbat was a great idea herrsheys
same but v2 is actually good
Its good for an open source model, at least it'd be good if it was gemma4's size
For qwen being a decently sized model its good its just not quite on consumer hardware yet
I have seen, its truly incredible
Kimi k 2.6
i agree
:D
why did arena didnt add kimi yet
Is 3.6+ better than max
idk i do think so
Announcement incoming π
Hello again
It did
they added only to battle mode then?
#1417174113092374689 message any possible ETA for a fix?
weird
wdym?
hey is muse spark not available in side by side??
at th ebottom
Lmao
sadly no
i now have to wait kimi 2.6 drop on nvidia nim api
nvidia nim api is the goat
free AI api
just 40 rpm
rate limiting
Pretty sure they added as placeholder for now and will ping when its actually on it
nah
they added it
Ah
i tested it
No clue why theres no ping then
I find datasets so intriguing
200k is very low context
Its standard
200k +
1M is high
also its standart
standard is 400k / 1m
actually good model
grok has 2m, gemini and claude have 1m, gpt has 1m / 400k depends on mode
im not joking
1M is high lol, highest we have currently is 2M
qwen has 1m even
Thats true
i just found my go to cheap opus level model
Ehh we will have to see
I feel like its just gonna be another GLM 5.1
its reasoning is going well
doesnt look like it
i used glm 5.1 and its HORRIBLE
i vomit when i see it
Glm 4.7 was better
Used it for a few tasks but when they released Glm 5.1, they made it worse and they also made Glm 4.7 bad π
weird my kilo code agent on kimi just stopped randomly
Honestly, it shouldnt even have to think THAT much to respond like it did
saying hi to gpt 5.4 pro extended is better (80$ lost in compute π₯ )
Mann I cant wait till my agent is trained, it'd be the first one I truly trained correctly
exactly likke claude
my code is a MESS
its horrible optimized because of glm 5.1
Ai just thinking to say an hi
im never touching a z ai model again
just use gpt
wow π
i wasted 2 dollars on that crap
Wait, what happened to the LLMs like ChatGPT 5.4, 5.5, Claude 4.7 Opus, and Metaβs newest proprietary model? Where did they suddenly disappear to?
Why have they been left out? Why canβt they be included in the comparison? This seems very suspicious and not quite right.
???
fr tho
They didn't put Opus
5.5 isnt even a real model
i wasted 6 dollars on it to fix the crappy code
LOL
In case Ampro says anything, ignore him.
π
buy the subscription bro not api
guys
im brazilian
im not wasting 40 dollars
on claude and chatgpt
me too βοΈ
I cant see them in the list of available llms in Side by Side
for me thats 200 reais
Recently some models have been removed from Direct and Side by Side mode. This was done to help ensure reliability and availability of Arena in the long term. You can find more information in this announcement: #announcements message
It's worth noting these models weren't removed from Battle.
is there a way to export my chat
I used codex 5.3 to build this agent and I just simply do the training
gpt models are getting great at coding
but i cant chat with them
they are horrible at chatting
wsp pineapple
thats why im in claude
Howdy 
but kimi is doing great for me]
For sure, I only use codex when it comes to GPT but they are getting real good
?
But Battle is RANDOM.It makes no sense.
which universe are you at?
sure it does
okay they are getting great
im gonna say
Yes to stop people from abusing and using opus 4.7 constantly then they run out of money
tho gpt 5.4 and 5.2 are good
the problem with claude is the pricing
Is it for ever, for good or is it provisional, temporary step?
if they fixed that
Brother what π Codex 5.3 is easily the best coding model out there
we would be rich
Not frontend but general coding its the best
Alright
i cant send this in chat
frontend its amazing
pineapple will kill me
i don't really like using codex claude code or something
not my type
It's our intent to bring the models back to Direct/Side by Side but will only do so in a way where it's sustainable.
Its not great frontend, backend is what its truly amazing at
Ah I get ya
Sorry send what? I'm in a bunch of chats rn 
one thing thats weird is that 5.4 is cheaper than sonnet so it makes no sense
what mode?
sounds rigged
but sometimes it give error
If you run into an issue sharing the problem in #1343291835845578853 with more of a breakdown on what happened would be helpful.
Bro just give me gpt image 2 model, I wanna make some neural slope images πππ
where
kimi response
its too big
Quite foggy and ambiguous as in new language π€£. Well, I hope you SHALL restore old settings. I do not believe that it will help Anthropic or GPT earn money. Arena.ai is for nerds like us.
I wouldn't mind it being shared
pineapple my brother drank a pineapple juice yesterday
Any chances you were involved in the drink pineapple juice crime?
I can not confirm, nor deny
Hm
π
okay then
fixes
he found
Here is the code review for v1 (nous/ and training/).
Critical Issues
1. Broken Causal Conv State During Autoregressive Generation
File: nous/generation.py (inference) / nous/model.py (SelectiveSSM.forward)
During generation, model() is called with a single token (next_token shape [1, 1]). In SelectiveSSM.forward, the new conv hidden state is computed as:
new_conv_hidden = x_proj.transpose(1, 2)[:, :, -(self.d_conv - 1) :]
When seq_len = 1 and d_conv = 4, slicing [-3:] on a length-1 tensor returns an empty tensor (shape [batch, d_inner, 0]). This empty tensor is passed back as working_memory for the next step. The 1D convolution's history is effectively reset every token, so the model loses local context across generation steps.
Fix: Maintain a fixed-size conv state buffer (length d_conv - 1) and shift-append new tokens, rather than slicing x_proj which only works when seq_len >= d_conv - 1.
Moderate Issues
2. deep_memory_init Buffer Is Never Used
File: nous/model.py lines 158β161
deep_memory = torch.zeros(
config.num_layers, config.batch_size, config.hidden_size
)
self.register_buffer("deep_memory_init", deep_memory)
forward() ignores this buffer and creates fresh zeros when deep_memory is None. It also hardcodes batch_size into a buffer shape, which is unnecessary. This is dead code.
Fix: Either remove it, or use it to initialize deep_memory in forward().
3. get_dominant_position() Hardcodes Slice Indices
File: nous/alignment.py lines 96β117
if alignment[256:384].norm() > alignment[384:].norm():
This assumes alignment_size >= 448. The default is 512, so it works out of the box, but if alignment_size is ever reduced below 448, this will raise an IndexError.
Fix: Derive slice sizes dynamically from alignment_size.
Minor Issues
4. ignore_index=0 Assumes Token 0 Is Padding
File: training/loss.py line 29
ignore_index=0,
This depends entirely on the SentencePiece tokenizer configuration. If ID 0 is a real vocabulary token (e.g., <unk> in some SPM setups), the model will never learn to predict it.
Fix: Make pad_token_id configurable in NOUSConfig and pass it through.
- Extremely Slow SSM Implementation
File: nous/model.py (_selective_scan)
The scan is a pure Python for loop over seq_len (up to 2048). On CUDA, this will be orders of magnitude slower than it should be. This is a performance bottleneck, not a correctness bug, but it makes v1 impractical to train at scale.
Fix: Use a vectorized scan kernel or consider a standard Mamba/CUDA implementation.
Design Observations (Not Bugs)
combined_loss phase weighting: Phase 2 uses 0.1 * next_token + 0.9 * coherence, which heavily suppresses language modeling. If this is intentional, fine β but it means the model may stop learning to predict tokens well during crystallization.
deep_memory injection: Every layer receives the same scalar-broadcasted final-layer mean (x.mean(dim=1)). If you intended layer-specific memory, this is not doing that.
alignment_update_interval = 10_000_000: This is so large that the alignment layer will likely never update during a standard training run.
Bottom line: The biggest problem is the broken conv state during generation, which will cause degraded/incoherent outputs. The dead deep_memory_init buffer and hardcoded alignment slices should also be cleaned up.
thats it
so gemini 3.1 pro, claude opus and etc still dont appear
And then you realize, it was you who was turned into pineapple juice.. And it's not pineapple talking, it's the MURDERER of the fruits..
bro
lol
lol got wrong server
server
server got wrong bro
bro got wrong bro
wrong got bro server
got wrong server bro
got bro server wrong?
server got bro wrong
bro got wrong lol
got server wrong bro
wrong did server get the bro out of my tongue
lol bro got server watermelon wrong
got wrong bro server
bro wrong server got
....
got wrong server bro