general

[loud herald]

OpenAI and Anthropic are the biggest scaling companies its insane

[wicked talon]

Bruh imagine what Google could actually do if they didn't optimise their ai models

[toxic verge]

Google can’t even make it to number 1

[wicked talon]

Imagine the data they have on users bro

[wicked talon]

They have

Not anymore after optimising

[loud herald]

Anthropic has very good pre training data, if they had the same amount of compute as OpenAI we'd have AGI by now

[wicked talon]

But imagine the data Google has it could probably build a better model then mythos but unfortunately it would cost too much

[loud herald]

Lmao imagine what they WILL do because they optimized their models to be smaller

And less compute

But just as strong

[wicked talon]

True

Cause they have petabytes of user data

[toxic verge]

OpenAI said it’s planning to have 30 gigawatts of compute by 2030, while it expects Anthropic to have roughly 7 to 8 gigawatts by the end of 2027.

[loud herald]

No it wouldn't, google has the most compute probably in the world but also out of openai and anthropic

[toxic verge]

Anthropic research

[loud herald]

Google has been around much longer thats why

[loud herald]

LOL

[wicked talon]

Google intentionally downscales their model for cost

[toxic verge]

TPUS suck

[wicked talon]

That's why they don't use top end models in all of its ai features.

[loud herald]

Because theres no point to going higher and higher it just gets more expensive

They see that they will get no profit from it

[loud herald]

So they optimize instead

[wicked talon]

Exactly

Mythos will never be released to the public

It will bankrupt anthropic

[loud herald]

Nope

[wicked talon]

Imagine the cost of it

[toxic verge]

The only reason not release is because they dint have the compute

[loud herald]

Mythos is $165 per million tokens

[wicked talon]

No because it was a risk

[loud herald]

No because they dont have the compute lmao

Their "its too strong!" was bs

[wicked talon]

Hmm might of been

[toxic verge]

Bro look

[wicked talon]

But they literally gave it to their competitors for free

[hazy fox]

Why are grok 4.20 reasoning and grok 4.20 multi agent so terrible, they write very bad and very short texts when I literally gave a detailed prompt where I wrote everything down

[toxic verge]

Now look what happens a year later

[wicked talon]

Does bro just have like every screenshot under the sun

[toxic verge]

What a coincidence

[wicked talon]

Hmm

[loud herald]

I'm attempting to build a model currently that is a sparse hybrid model TIFSM I have 6GB of vram and using methods from bitnet to build with binary integers and it uses MUCH less VRAM than a transformer. I will be training an 8B Dimension TIFSM and it will use 4gigs to train, while having a context and memory and general intelligence of a 30B+ Parameter transformer

Its theoretical currently

But if it works out after training

It is 100% true

[toxic verge]

They’re only releasing it to the big corporations and big companies that have money that they already have business with their integrated into some sort of business with them

That they had business relationships with prior to mythos lol

Because they don’t have the compute to release it to the general public

Not yet at least

[loud herald]

This is the "wall" people talk about

Anthropic needs to go public

[cedar citrus]

hey, it's davitotty — i've actually been running sparse hybrid experiments (SSM + attention + micro-experts) and the VRAM math just doesn't work the way you're describing. BitNet cuts weight memory, sure, but gradients + optimizer states + activations during training will blow past 4GB on anything 8B-scale. also 'general intelligence of a 30B+ transformer' from a compressed model isn't something you can claim before running a single benchmark. cool direction though, prove it with evals and i'll believe it

[loud herald]

I never said I claimed it, thats why I said I'm attempting to build it and said its theoretical afterwards

[hazy fox]

Who knows?

[cedar citrus]

fair, missed the theoretical part — curious to see the evals when you get there though, the VRAM claims are the part i'd stress test first

[toxic verge]

Once they get locked in, you’re gonna be paying premium prices for these top models

X2 x5 times more

[loud herald]

Yea thats what I've been seeing, but once again it matters not how big it is but the active sparse with how much VRAM it uses

[cedar citrus]

yeah fair, active params is what matters at inference — if the sparsity routing is efficient enough the 4GB claim becomes a lot more believable

[cedar citrus]

what does TIFSM stand for btw

[loud herald]

True that, but we will see when I train this thing lol. If my PC explodes I will let you know

[cedar citrus]

k send me the results

ill love to see it

[toxic verge]

We are the guinea pigs for the consumer grade ai

[cedar citrus]

im also trying to do similar experiments

[loud herald]

Topographical Invariant Functional Sparse Mapping, believe me, I didn't come up with the name

[toxic verge]

I heard something about that recently didn’t pay no attention to it though

[cedar citrus]

can i dm you

im curious to see your idea

[loud herald]

Sure you can

[tired herald]

how incredibly interesting

unfortunately you will face a massive road block

actually multiple

First is high quality dataset that you can actually reliably use

and second is the compute required to build multiple billion parameter models

[wicked talon]

Wait why does grok have a therapist mode 😭

Low-key I wonder how it's like

[loud herald]

Yes thats true, I need to get a high quality one

[tired herald]

@loud herald with your device you are looking at training 100-200m models at incredibly slow speeds with small context windows for tens of hours

[loud herald]

See thats why this is different because it doesnt use it like a normal transformer

[tired herald]

could you elaborate?

[loud herald]

No I know that, if it were a transformer but tis not, I could train 250M param max with my 1660 on a transformer a TIFSM cannot

[tired herald]

I personally have tried some stuff and it just takes a long time

[loud herald]

I'd like to send a text file explaining it if you don't mind, this should get what I cant get out because I am really bad at explaining stuff

[tired herald]

go right ahead

[still musk]

😍

[loud herald]

Indeed it does but in this stuff you gotta be patient

[tired herald]

i have some exp in this since ive tried some stuff

[median smelt]

Whats wild-bits?

[tired herald]

lemme read through for a sec

[loud herald]

It may not explain well actually but any questions I could probably answer after

[wicked talon]

How has bro pulled this off

[median smelt]

HOLY MOLY I GOT A/B TESTING

boii

[wicked talon]

Eww chatgpt

[median smelt]

Im gonna chat on ur gpt

[tired herald]

well

to say the least

[wicked talon]

Bruh last time deepseek and Claude had a battle Claude cried

[tired herald]

the idea is kinda passable... Its very close to the idea behind MOE models

[loud herald]

I'm gonna be honest, it has about a 10% chance of it really working how I want it to work so its experimental

[sterile storm]

doesnt work bro

[tired herald]

What?

[tired herald]

No, we're talking about 1-2% chance

[wicked talon]

Ahh that's how grok gets away with therapist mode

It's in quotes

[loud herald]

I am implementing MoE into my 8B Dimension model so it will have that in there, but MoE also uses a lot of VRAM so I think what I will do is whichever MoE isnt active will be dormant on my ram until it needs used

[loud herald]

Same same

All insanely low

[tired herald]

Whatever the case, allow me a second to try something

[loud herald]

Sure

[tired herald]

I can do some testing for your idea here

[quaint trail]

anthropic is the most ass company holy

[tired herald]

I personally have been playing around for shits and giggles and will quickly implement in the rough TIFSM idea

[loud herald]

Alright

[median smelt]

Guys

What is wild-bits

[loud herald]

I have no clue it sounds interesting

[inland ember]

Readd duct!

All of the image models are ass

[zealous sparrow]

when image models get removed they are nearing release

[inland ember]

possibly becoming
gpt-image-2 standard
gpt-image-2 middle
gpt-image 2 high

[grave peak]

https://tenor.com/view/epstein-jeffrey-epstein-j-jeffrey-sigma-gif-405083941180536765

[uneven cypress]

I just need Claude opus 4.7 search mode

[tired herald]

they dont even have that normally 😭

[grave peak]

yes we need this boi

[uneven cypress]

Ik I want them to add it🥱

[median smelt]

we need image 2 bou

Boii

[tired herald]

sometime in the next 10 years TM

[uneven cypress]

Wait how do u have it in direct chat

[tired herald]

Its a magic trick 😭

[uneven cypress]

Tell me🙂

[wicked talon]

Kimi rated gpt 4.1 over mythos lmaoo

[tired herald]

I modify website code before it loads to change the settings of all available models to allow userSelectable

[uneven cypress]

And the models actually works?🤔

[median smelt]

no

Because dudbddhbdbc ftbb g c cjf dcchcnfdbccc

[half yew]

Can i get tut in dms

[tired herald]

no because they disabled that server-side

[uneven cypress]

Ok

[tired herald]

No, first, because I would be banned, second, because its useless

[still musk]

OMG

NEXT MONDAY GPT IMAGE 2??

[median smelt]

Maybe?

[still musk]

Next month????

[median smelt]

Prob earlier than next month

[remote dove]

That's odd

[still musk]

Let's hope so

[remote dove]

I thought opus models wasn't on direct?

[night moat]

Seedance 2 when?

[tired herald]

prob because opus 4.7 currently isnt thought to be used by the public so yeah

[grave peak]

@echo aurora

https://giphy.com/gifs/meme-tuff-say-wallahi-T0evwqyUZsqNEJzmbW

it s a hope

[icy remnant]

hwo to know anthropic removed from arena or rate limit hit of mine accoiunt for use?

[loud herald]

It will tell you

[icy remnant]

who?

[half yew]

Here’s a result i got frm duct tape

[median smelt]

Result from chatgpt

[grave peak]

what is this?

[median smelt]

Image 2 on ChatGPT A/B Testing

[lofty frigate]

This model is soooooo good

[uneven cypress]

Claude from temu

[tired herald]

Temu would've made better AIs than hofburg

[gray lion]

claude literally got a simple biology question wrong idk whats happening, the other model in battle got it right

[wicked talon]

Claude is just bad lol

[gray lion]

normally its good but maybe this is just 4.7 bc 4.6 always was good

[wicked talon]

Claude 4.7 so censored it would refuse to write a breakup text tbf

[molten robin]

the UK government

[wicked talon]

Mullvad Vpn + grapheneos comes into play

LMAOOOOO

[vagrant temple]

Has anyone else verified at Gpt Cyber?

[frank junco]

https://giphy.com/gifs/hypergoku7-db-meme-vegito-roster-qqVtsdZcYsTsjvQqwq

[molten robin]

i live in the UK so i have to use a VPN to go ONTO IMGUR

[surreal zephyr]

😍

[tired herald]

Training: 1%|▍ | 118/10000 [00:40<59:25, 2.77step/s, loss=8.348, lr=2.4e-04, ep=1] await

its a small af model but its enough for some minor testing against other techniques

[loud herald]

Noice! Also you were saying earlier about how running an 8B dimension model on TIFSM would run my GPU into the ground, but I believe you may have been mistaken

As you know, sparse models aren't dense and dont use all the weights for each token prediction

But I think we did go over that earlier

That its all about how you tune the active sparses

Because it may sound insane

But

If you tune and train a model a certain way an 11 gig VRAM card could run an 120B dimension model

At least theoretically

But unbelievable tuning would need to be there lol

How so lol

[tired herald]

Im not sure what you are talking about here, training a model from scratch on TIFSM is mathematically impossible to train an 8b on 6gb of VRAM.
Are you possibly talking about fine-tuning a model into your arch or what? (ignore dis stupid ass idea)

man my english is so dead rn

or maybe training a sparse autoencoder?

[golden ocean]

https://tenor.com/view/dog-cycling-police-police-chase-bike-gif-13680985422962193768

[tired herald]

nice doggo

[loud herald]

Because it uses a Fixed Active-K of 8192 the mathematical operations are limited to a tiny fraction of the total space at any given time, By using Expert Paging, I’m keeping the dormant experts in System RAM and only 'activating' the relevant sparse indices in the 6GB VRAM

Its the same as a sparse model just using specific weights for a specific question

[tired herald]

well

[rose jackal]

4.7 is downgrade

4.6 was much better

and still better

[tired herald]

im pretty sure thats just moe-ish sparse indexing 😭

[rose jackal]

but anthropic nerfed it

[loud herald]

This has integers for specific "agents" inside that speak to each other and theres multiple of these sequences, and then theres a director that decides which one is best for the answer

[loud herald]

Honestly it probably is

But it uses much less vram

[tired herald]

and that is basically cpu offloading explained in a more words

Gimme a sec

[loud herald]

The whole thing behind doing this is to get the best quality AI out of the smallest of VRAM I can do

[tired herald]

Then allow me to give you some criticism

[loud herald]

I am completely open to it

[tired herald]

What you're doing is called MoE.
Everything you've explained already exists, and has been heavily optimized by professionals.
You are trying to "reinvent" what already exists and are giving yourself a massive headache.
I would simply recommend going with a true MoE model and thus save yourself the trouble of fixing issues with your TIFSM.

[toxic verge]

https://tenor.com/view/omg-popcorn-gif-3983484043797832117

[tired herald]

😭

If its not broken then dont fix it type shi

[loud herald]

I'm not trying to reinvent anything, I said before it a hybrid of multiple different things

But yea MoE is a big part in it

The whole purpose of trying this like I said was trying to get the most out of little VRAM I have and if it works then it works

If it doesn't then I will move on

[tired herald]

Well, thats what it looks like to me, since you think that, i wholly believe you.
Go ahead and do your thing and I hope it works for you, but do remember that you shouldnt go expecting any results in any short amount of time.
Good luck!

[loud herald]

Yea I agree, I dont get my hopes up about this stuff but I will keep trying, and thank you

[tired herald]

(btw, I heavily recommend not trying to train anything from scratch and simply using the best of the best and "shaving" them off, whatever you think that could mean tweet_glock )

[loud herald]

I have no idea what you mean there 🤣 but you may be right, even so I take joy in creating my own models

From scratch

[tired herald]

😸

Well i have a basic 40m model that is somewhat coherent so I do understand

I recommend using OpenWebText, FineWeb, WikiText for actual training (base model training) and if you want to quickly test something out, use TinyStories

[loud herald]

I use FineWeb-edu and recently used WikiText, FineWeb has always worked out for me

[tired herald]

Oh no!

thats not good to do

[loud herald]

Why?

[tired herald]

FineWeb should not be used alone because it really isnt good for basic Training

[loud herald]

Ohhhh

[tired herald]

You should take data from multiple datasets to diversify data

[loud herald]

I see, that makes sense idk how I never thought about that

[tired herald]

You'll run into the problem of the model only doing stuff that is available in FineWeb, and FineWeb really f*cked up my early model training because my models kept giving bad outputs

[loud herald]

Oh boy

[tired herald]

but when i changed over to using multiple datasets and limiting FineWeb to like 20% of my training data, i saw massive improvements

[loud herald]

I am using multiple datasets for this hybrid this time, I dont think I've ever trained on so many lines LOL

[tired herald]

good good

Are you keeping the Chinchilla rules in line?

of 20 tokens per parameter for training data

[loud herald]

I didn't know about that

[tired herald]

You need that much training data min if you want any good quality

[loud herald]

I'm gonna be honest I've only been training for about a month and a half so I dont know too much about it currently

[loud herald]

Okay

[tired herald]

I've been training for like 1.5 weeks 😭

[loud herald]

Jeez laweez

[tired herald]

Well, I have been RE'ing for the past 2 years so maybe that could be the reason

[loud herald]

Ahh

Possibly but still

Good job

[tired herald]

second

    [temp=0.5]: Tokyo is the world's most important city, and it has been home to a number of high - profile cities over the past decade. But one thing is clear: Japan is not just an outlier in its own right. The country is  (50 tokens, 13 tok/s, 3758ms)
    [temp=0.8]: Tokyo is having to take a decision on what happens next. The Japanese government has agreed to move the project from Ryokuganshi in northern Japan, and will do so within six months of its completion. A team of about  (50 tokens, 10 tok/s, 4764ms)
    [temp=1.0]: Tokyo is the biggest power - efficient city in the world, and its population is on par with those of New York City. There are already a large number of places like this that have no electricity production or electricity source to store or supply power that requires an industrial  (50 tokens, 13 tok/s, 3934ms)
``` my 40m model

and it hasnt been trained on that much data yet

[loud herald]

Thats crazy

[sturdy siren]

A realistic fitness advertisement scene in a modern living room with warm natural sunlight coming from a window.

A confident woman in her 30s is doing a plank exercise on a yoga mat, wearing simple grey workout clothes, focused and strong.

Place a realistic ebook next to her on the floor with a feminine pink design cover titled "The Hormonal Belly Reset Guide (21-Day System for Women 35+)".

Cinematic lighting, soft shadows, depth of field, premium lifestyle photography.

Add clean white modern text overlay:

Top text: "Hormonal reset for women 35+"

Left text: "7-minute routine"

Right text: "daily"

Small text: "Simple system that fits your daily life"

Bottom text: "Flatten your belly, boost your energy, and feel confident again"

Add thin curved arrows connecting the text in a minimal infographic style.

Slight slow zoom camera movement, Instagram ad style, clean and high-end look.

[tired herald]

    [temp=0.5]: My life has been a mess. I have to go back to my old house and try to get some sleep."The day before his death, he was in the hospital with a heart attack and had not been able to return home for two days. He  (50 tokens, 13 tok/s, 3748ms)
    [temp=0.8]: My life is an absolute mess, but I have never been able to be a fan of the'80s. It's just like any other college town where the backside Bronco was born. This school is  (54 tokens, 14 tok/s, 3791ms)
    [temp=1.0]: My life is about to go dark and I have not given a single thought on it. But there is still hope in me, there are others with whom the world knows its going to turn against me. One thing that should only be done when I have no  (50 tokens, 13 tok/s, 3820ms)``` Darkness myyyy frieeennd

My model kinda seems very negative 💀

[loud herald]

LOLL

Its the depression cycle

[tired herald]

    [temp=0.5]: Tokyo is in the midst of a major crisis, with an estimated 2. 5 million people living on their homes and over 1. 6 million children suffering from injuries. The government has said it will not accept any new measures to help the poor, but will also consider  (50 tokens, 13 tok/s, 3724ms)
    [temp=0.8]: Tokyo is in the midst of a rout of financial troubles. Some of its biggest and most valuable assets are set to be invested by the Japanese central bank. Some of them have been under pressure from European banks to fund foreign investment, such as Ki  (49 tokens, 13 tok/s, 3632ms)
    [temp=1.0]: Tokyo is in the process of forming a new team to represent Japan, and has already been through an open meeting with members from Europe (Asia, Pacific, Southeast Asia, South Asia, Southeast Asia). At this point, it seems like we're not at all sure  (50 tokens, 13 tok/s, 3729ms)```

[loud herald]

You gave it the wrong side of the internet

[tired herald]

😭

Wait

I have something interesting I made to help train larger models....

[light sleet]

guys what news should I add to arrna

arena news

website

Any article ideas? (your name and pfp would be there)

[tired herald]

"A potential Credit System incoming: Is this the beginning of the end""

[light sleet]

ok do u want to be the publisher (even tho I'll add article with ur name and pfp)

[tired herald]

Yes?

[light sleet]

ok

DeOld779

[loud herald]

Oooo

[vale quest]

Nah that's north korea

[tired herald]

I call it surgery

[light sleet]

gimme ur pfp in dm

[tired herald]

second

[loud herald]

Hows it work

[loud herald]

https://vaultcord.com/tools/discord-id-lookup

[vale quest]

Or use

Bluecord

[loud herald]

Either is fine

I just clicked the first one that popped up

[vale quest]

Ngl that one sucka

Sucks

Use pfpfinder.com

[tired herald]

or just use f12 in chrome discord to extract the image

[tired herald]

well, it basically takes existing parameters, does some fancy shmancy math and creates new parameters. The new model would fully work, and you wouldnt have to train the model that long to get it to the same point as training it brute force

[loud herald]

Ah I see that’s pretty awesome

[tired herald]

yeah, thats how i got a 40m model to 120m (and with a bit more training) got to really good performance

at least good for its size

[light sleet]

nvm I probably can't credit u

pineapple I can tho

cuz

I could take his possibilities

reasons

[tired herald]

just credit my discord @ at the end

[cedar citrus]

guys

im right now training ai for philosophy actually im trying to install mamba ssm

super annoying process but lets me train faster

[tired herald]

nice

are you fine tuning or training from scratch?

[light sleet]

Nvm I should prob post the news from announcements when they come

[tired herald]

🦥

[cedar citrus]

scratch

im training it on philosphy

and seeing on what it alignes with

its a RNN

[tired herald]

😭

ohhh

[cedar citrus]

it will completely generate the text so no LLm to help its only the neural network

2b paramaters

it will be

[tired herald]

Damn

[cedar citrus]

im training two versions

a v1 which is a normal rnn

and a v2 my version

[tired herald]

what hardware are you using

[cedar citrus]

my dad has a 4070

so im using his pc and he has a amd ryzen and 64gb

ram

gonna take 10 days

im having to use wsl2

mamba smm isnt suitable for windows

and mamba smm is a pain to install

[tired herald]

How is it an RNN but also a Mamba SMM 😭

[cedar citrus]

k wait

ill answer

[tired herald]

take your time

[cedar citrus]

PHILOSOPHOS is Mamba SSM — not a classical RNN, though Mamba is in the same family conceptually.
The distinction:
Classical RNN (LSTM, GRU)

Sequential hidden state, gated updates
Slow to train (can't parallelize across time)
Struggles with very long-range dependencies

Mamba (SSM — State Space Model)

Also maintains a recurrent hidden state
But can be computed in parallel during training (like a transformer) via a scan algorithm
Selective state spaces — learns which input to let into the state, which is the key innovation
Much faster than classical RNNs, competitive with transformers on long sequences

PHILOSOPHOS/NOUS specifically

V1 is the base Mamba SSM architecture trained on philosophy text
V2 (NOUS) adds the MoE hybrid on top — Gumbel/TopK expert routing layered into the SSM, which is where your salience_gate.py lives
The "philosophy RNN" framing you used earlier was just a loose description — the actual backbone is Mamba

So the short answer: Mamba, which behaves like an RNN at inference time but trains like a parallel model. The confusion is understandable because Mamba is often described both ways depending on context.

[light sleet]

https://019da287-8bbc-7f70-a54c-2487091624ab.arena.site/

[cedar citrus]

tecnically

its a RNN

mamba is a type of rnn

[light sleet]

New Article added (old article but it's about opus 4.7 being added)

[cedar citrus]

how do you get a job at lm arena lol

[tired herald]

Careful! Mamba is a further improved version of what RNNs are. Its not an RNN!

[light sleet]

I have no job 😔

[tired herald]

but that doesnt matter

[cedar citrus]

Mamba IS a type of RNN. Specifically, it's a Selective State Space Model (SSM) — which is a recurrent architecture that processes tokens one-at-a-time using a hidden state, just like any RNN. Your code proves it:

[tired herald]

💀

[cedar citrus]

its a TYPE of RNN

[tired herald]

But that doesnt matter.

I simply wish you luck with your training

Just dont expect any results comparable to existing models of similar sizes by large labs

[tired herald]

oh no

[faint willow]

what is this

[cedar citrus]

the objective is just seeing what ais aligne in philosophy

its not general knowledge

[toxic verge]

https://www.reddit.com/r/GeminiAI/comments/1sosxi2/no_more_than_1_image_in_latest_update/

Anybody else experiencing this?

[tired herald]

Well, the model is gonna align with whatever exists most in the training data (is what i think would happen?)

[cedar citrus]

its gonna be made specifically for this

and it has the own thingies

[void shore]

Im gonna be making my own ai model as well, because i have a huggingface account:
https://huggingface.co/GenueAI (check it out if you can)

[cedar citrus]

not out for me yet

[light sleet]

I was jk

not out for me too

I meant its out for Sam Altman

Cuz hes the ceo and he probably gets access to dev betas

[void shore]

Im working on a new style for datasets, where i write seeds:

User: I have this pattern [q], what is the next item in the sequence?
Bot: Lets follow the pattern, it does [q], figuring out the solution is [s], the answer is [a]

Then fill in the patterns

like 50 times as new examples

and change the sentence a bit

and i have 25 examples to like 1500 examples about

because i dont like using other datasets

because i like to have control over it

if theres an issue with the bot

[hollow ivy]

-# (Afaik, GPT does not have the longest context, therefore its no option in this poll.)

[polar horizon]

hydrogen bomb versus coughing babies

bro chatgpt's PR is mad because he's not even top 5 in anything text related

[vital lake]

Obviously mimo

[acoustic garden]

where did the opus go?

[tired herald]

Massive Overfitting issue

[void shore]

i am adding a lot of variation between prompts, like this for example:

{"user": "If it's cold outside, wear a jacket, if it's hot outside, wear shorts, currently it's cold outside, what occurs?", "bot": "To calculate what will happen, I would check that it's cold outside, so wear a jacket will happen, we get wear a jacket"}
{"user": "If you're hungry, eat food, if you're thirsty, drink water, currently you're hungry, what's the result?", "bot": "To calculate what will happen, I would check that you're hungry, so eat food will happen, the result is eat food"}
{"user": "If the light is green, go, if the light is red, stop, currently the light is green, what's the result?", "bot": "To calculate what will happen, I would check that the light is green, so go will happen, the result is go"}
{"user": "If it's cold outside, wear a jacket, if it's hot outside, wear shorts, currently it's cold outside, what's the result?", "bot": "To determine what will happen, I would check that it's cold outside, so wear a jacket will happen, the result is wear a jacket"}
{"user": "If the light is green, go, if the light is red, stop, currently the light is green, what's the result?", "bot": "To calculate what will happen, I would check that the light is green, so go will happen, the result is go"}

If theres any changes

i would like to know

[hollow ivy]

i'm not only interested who wins this poll, but also who's the runner-up

[void shore]

the grammar is bad, i know

[tired herald]

again

[clear spear]

I'm so pissed I know arena can't handle the model costs but I really wanna use it and there's nowhere else that offers them!!! at this point someone pay for my stuff! I have cashapp 😭

[void shore]

im gonna do everything i can to make sure it wont overfit

like noise, punctuation

[tired herald]

massive overfitting issue.
data should be diverse, as in, different sentences with different styles, but ones that maybe say the same thing

gimme a sec

bruh

is what im talking about

[void shore]

Thanks!

[tired herald]

Your data would kill the possible performance of a model

Making the data diverse does actually heavily improve performance across the board

[void shore]

thanks for telling me this

im updating my dataset expander script

[tired herald]

Always happy to help

My ass tried so much stuff over the past 2 weeks of training AIs on my own...
But i made mistakes such as the ones you got into and wasted so much time.
Next time you have an Idea, simply try asking your nearest AI for help (iykyk) and most of the time, they will tell you if there are any problems with your idea or how it could work

[still musk]

OMG

[tired herald]

Polymarket is almost never even close to being correct with this stuff

[still musk]

Well, if he said it's confirmed for next week, we'll see, right?

[tired herald]

I hope it is coming next week lol

[still musk]

MAYBE, I say MAYBE, in addition to GPT 5.5, GPT image 2 will be released

But it MUST come out

[toxic verge]

The arena is is broken 😡

[still musk]

GPT IMAGE 1.5

[loud herald]

So this coding dataset I have is mixed up of a ton of different coding datasets and also distilled model datasets and its 10 Million lines, its going to take an eternity to do LOL but do you think thats smart?

[glacial dock]

It is now time for the joke of the day

They say you should never go to Starbucks sick…
Or else you'll get a …cough fee !

thank you

[loud herald]

Wow

Awesome joke

Good job

[waxen grove]

why did the duck feel discriminated against?

the swan told him to duck off

[glacial dock]

What duh quack

[still musk]

However the alternative for GPT Image 2 for now if you want to use it on Arena AI is Nano Banana Pro by Gemini 3

[void shore]

I am starting training, i have made sure my dataset is free of duplications

[still musk]

GAHAHAAHA

[void shore]

it went down to only 200 good examples

[pallid crypt]

Is that really enough data

[void shore]

its not from scratch

im fine tuning qwen 2.5 to test

[pallid crypt]

Oh ok

[void shore]

i need to see if my dataset even works

[still musk]

AHAHAHAH

[tired herald]

Good good

[tired herald]

Even better

[tired herald]

Yes! Its really good to have different datasets, even if its all the same overall type of dataset.
Unfortunately I cant tell you if 10 million lines is good since i cant exactly say how much data that is, but it sounds like it should be enough to train, or fine tune a tiny model

[thick pawn]

Couldn't be any more true 😆

[light sleet]

guys

@still musk

new arena news update in a few mins

[still musk]

👀👀

[light sleet]

Introducing Chatroom (talk to other people live),
With AI

Named Newsy

U can use AI in Chatroom.

Tag @Newsy to talk to him in Chatroom (test Chatroom w me pls I wanna see if it works)

[still musk]

Ah

[loud herald]

Very cool

[still musk]

Very interesting

[undone saffron]

Ne‌w Up‌da‌‌te -

@‌‌Ar‌‌e‌n‌‌a U‌‌p‌‌da‌‌te‌s - So‌me chan‌‌ges were ma‌‌de to the pl‌‌at‌fo‌rm in Te‌xt Are‌na, Video Arena and Co‌de Are‌na!

• The limit has been reduced to 5 messages daily
• The character limit has been reduced to 10,000 per message
• You can only have 15 chats per account; to create new chats, you must archive some
• You can only generate 2 videos and 1 image per day
• Chats are available for 1 month, after which they will be automatically archived

[tired herald]

So real

[undone saffron]

Yes

[tired herald]

Arena(Yupp2.0) AI™

Jk

[light sleet]

arena news here

https://019da2e7-9bc8-7156-a4dc-7a665858de2a.arena.site/

someone go to Chatroom and start chatting

Let's hope it works

[undone saffron]

wss://broker.hivemq.com:8884/mqtt
sus

[tired herald]

Dead

[light sleet]

I replied

u said fake

[tired herald]

Gmmm

[light sleet]

Who am I talking to in Chatroom

[undone saffron]

That wws don't work for me

[vernal zodiac]

hey guys any news about duct tape?

[undone saffron]

No

[light sleet]

Who's this 😭

[undone saffron]

Claude sonnet

[light sleet]

yo manifest v2 girl come in the Chatroom in arena news

[tired herald]

Funny thing

[undone saffron]

I did it 3 mins before
And don't work for me

[light sleet]

Bro try in different browser

when did I type bro

Remove bro

Autocorrect

[tired herald]

In my head i heard you saying bro in an indian accent

This is never gonna leave my head now

[light sleet]

😭

[undone saffron]

#general message

[light sleet]

manifest v2 girl finds old text quick

New AI?

manifestv2girl-websearch

[tired herald]

😭

[light sleet]

if pineapple was alive today he'd be shocked

😔

but I think he in like a break

Let em rest too

[tired herald]

I go to sleep

[light sleet]

nice

bye

goodnight

[tired herald]

Gn my brethren

[vernal zodiac]

🫣

[light sleet]

3:33am for me

🔥 🔥

[light sleet]

Nah

Still same and VERY rare

Although the tracker said it's removed

Sevatar come to Chatroom https://019da2e7-9bc8-7156-a4dc-7a665858de2a.arena.site/

Pull up in the Chatroom

[vernal zodiac]

ah that's rough

[vernal zodiac]

sorry bro i am gonna hit the sack

it's 2:35

[light sleet]

FAHHHHHH

Ima get off Chatroom now

[vernal zodiac]

😫

damn i noticed that duct tape and the new ai model (a/b) in the official chatgpt are totally different

[whole sundial]

you are already only allowed to make 2 videos per day btw, should be changed if you post this again

[light sleet]

PLS KIRI

I NEVER TALKED TO U

PLS

Come to Chatroom in Arena News

https://019da2e7-9bc8-7156-a4dc-7a665858de2a.arena.site/

PLS 😭 🥀

Kiri I saw ur text

[still musk]

WHAT? 1 IMAGE A DAY? WHAT DO YOU MEAN?

[light sleet]

u wrote test

[undone saffron]

Yes, it's old -> #general message
-# Fun fact: If you get an error when sending the prompt to generate the videos, it still counts as a generated video, and you have to wait 24 hours.

[light sleet]

manifest v2 girl what's ur real name

Or ima js call u manifest v2 girl

[undone saffron]

Of course I'll say that here

[vernal zodiac]

what

[light sleet]

I meant not irl name

like a discord name

like I'm banana

[undone saffron]

Or "wolfy" like on twitch

[light sleet]

kk

I hopped off Chatroom

[whole sundial]

yeah this is fully gone from arena now, there are no tape models at all

[light sleet]

arena news

🔥 🔥

I'm js waiting for pineapple to see it

😭 😭

[undone saffron]

Ne‌w Up‌da‌‌te -

@‌‌Ar‌‌e‌n‌‌a U‌‌p‌‌da‌‌te‌s - So‌me chan‌‌ges were ma‌‌de to the pl‌‌at‌fo‌rm in Te‌xt Are‌na, Video Arena and Co‌de Are‌na!

• The limit has been reduced to 1 messages daily
• The character limit has been reduced to 5000 per message
• You can only have 5 chats per account; to create new chats, you must archive some
• You can only generate 1 videos and 1 image every 15 days
• Chats are available for 1 week, after which they will be automatically archived

[rigid pasture]

what is happening?

[light sleet]

I made a website called Arena News using Arena

basically all articles are from the announcements channel

It now has Chatroom and an AI named Newsy

[cursive cape]

What

[raven otter]

Now unusable!

[vernal zodiac]

is it unknown whether they will bring it back in the future?

[undone saffron]

That's sarcasm

[vernal zodiac]

or like maybe they are removing it entirely before the release

[light sleet]

Won't

Removed entirely 100%

[whole sundial]

it may or may not come back, the model will be launched next week anyways

so probably not

[light sleet]

Cuz it's releasing this Thursday or might be this Tuesday too.

[raven otter]

NEAT! That means I can make more of THIS

[still musk]

It won't come out, trust me, I've lost hope now😭

[whole sundial]

i just tried to make 2 images in chatgpt (idk what the free daily limit is) and both are image 1.5

[vernal zodiac]

wow next weak

is that for real?

[still musk]

GPT image 2 will NOT be released on Arena AI

[still musk]

It will be released in 2028 hahahah

[vernal zodiac]

btw i think it will drop on 23 april thursday

[light sleet]

hehehhe

I love being a produce

[vernal zodiac]

or sometime between april 23 and may 14

[raven otter]

New channel?

[light sleet]

it's like having a backstage pass to talk to Pineapple

[light sleet]

yes but only for us produces

u can be a produce too but u gotta join us and change ur pfp

To a fruit like Pineapple's.

Or anything.

[raven otter]

Don’t get our hopes up man

[whole sundial]

ok but you do have these

[light sleet]

well are they alive

[still musk]

I'm waiting for GPT Image 2 like the rise of Christ

[whole sundial]

yeah

[light sleet]

produce lot too

[undone saffron]

Ne‌w Up‌da‌‌te -

@‌‌Ar‌‌e‌n‌‌a U‌‌p‌‌da‌‌te‌s - So‌me chan‌‌ges were ma‌‌de to the pl‌‌at‌fo‌rm in Te‌xt Are‌na, Video Arena and Co‌de Are‌na!

• The limit has been reduced to 0 messages daily
• The character limit has been reduced to 1 per message
• You can only have 0 chats per account; to create new chats, you must archive some
• You can't generate videos and images anymore
• Chats are available for 0 days, after which they will be automatically archived

[rigid vine]

[still musk]

💀

[cursive cape]

https://tenor.com/view/mengsedih-gif-23325538

[cursive cape]

@‌‌Ar‌‌e‌n‌‌a U‌‌p‌‌da‌‌te‌s - So‌me chan‌‌ges were ma‌‌de to the pl‌‌at‌fo‌rm in Te‌xt Are‌na, Video Arena and Co‌de Are‌na!
The limit has been reduced to -1 messages daily
The character limit has been reduced to 0.5 per message
You can only have -1 chats per account; to create new chats pay $100
You can't generate videos and images anymore (pay $100)
Chats are available for never, after which they will be automatically archived

[fiery gull]

no qwen3.6 35b in lmarena ;-;

it´s soo good this model

[cursive cape]

._.

[warped lagoon]

Anybody gonna explain why lmarena literally has Claude opus 4.7 in the leaderboards and it cant even be used... lol

[cursive cape]

it can only be used in the "2 random models" mode or something like that

[fiery gull]

test the qwen3.6 35b

better that qwen 3.5 plus

[cursive cape]

3.6 plus better

[fiery gull]

35b>>>

same thing, more fast

even in code

[cursive cape]

don't know

[warped lagoon]

Only 3.6 plus is on lmarena tho lol

[cursive cape]

I'm waiting for the deepseek v4 or better opus 4.6 model to be released so I can leave platforms like this one, but it's still good

[fiery gull]

have the aistudio free and qwen

[faint willow]

is duct-tape even on arena anymore

[cursive cape]

They're still hallucinating

[fiery gull]

the free limits of 3.1 pro are good, you can use multiple accounts

[cursive cape]

They're still hallucinating bro

[fiery gull]

ah okay, now I read correct

[cursive cape]

What do you think will be close to the level of opus 4.6?

https://tenor.com/view/shy-embarrassed-nervous-avoid-uncomfortable-gif-9785329523072125399

[fiery gull]

Idk, I think the deepssek v4 will be the most cost-intelligence model, I tried to use the minimax 2.7 but it is soo buggy, ever time it send chineses tokens, idk is more broken now, my english or m2.7, I'll sleep, good night

[cursive cape]

good night

[fiery gull]

From what I've tested, Deepseek is the model that most closely like claude's model (aside from raw intelligence), which is a good thing, but it seems Antropic didn't like it, lol

[cursive cape]

Yes, I think she uses something like leaks for training

I've heard it somewhere

[fiery gull]

I think the competition will be between the Sonnet 4.8 and v4, but it seems Deepseek will only release the v4 when it's the "big deal" like the GLM 5 was (they kept delaying the number 5 with 4.6 4.7...)

release the NAME v4

[cursive cape]

perhaps

[fiery gull]

if is just good model they will release the name v3.3, idk

[cursive cape]

The release was promised for the end of April, and I hope it happens

GLM is just terrible, I think the level is lower than sonnet 4.5?

or 4.6 haikyu

[fiery gull]

Yes, I find 3.2v a bit dumb for my tasks, but its personality is very good (important for my work) because it's copied... I mean, similar to claude's

[void shore]

i have finished my first run! Results:

User: What is 15 multiplied by 4?
Assistant: To calculate 15 times 4, I'd multiply 15 by 4, the result's 60, we get 60, the answer is 60, let's do it again, the output is 60, so it works, the solution is 60, here's the process: 15 * 4 = 60, that's the answer.

==================================================
User: All humans are mortal. Socrates is a human. Therefore, what is Socrates?
Assistant: To calculate that Socrates is mortal, since all humans are mortal and Socrates is a human, the conclusion is that Socrates is mortal.

==================================================
User: Describe a sunset in one short sentence.
Assistant: The sun sets, casting a beautiful orange and pink glow across the sky.

==================================================
User: If it is 3pm and I wait 2 hours, what time is it? Then tell me if it is late or early.
Assistant: To calculate the new time, I'd add 2 to 3, we get 5, so it's 5pm, that's not late, it's early!

the only one it messed up on was 15 x 4

it kinda rambled

[fiery gull]

I say personality, but it's not quite that, it's more about response style, like Flash 3.0, he's very lazy in his responses and also uses more vulgar language

[cursive cape]

AI? How do you train it?

[void shore]

Using Pytorch, Qwen, and a good dataset

[cursive cape]

We are waiting for the model on arena.ai 😂 👍

[void shore]

its releasing dont worry

just need to get an inference provider

anyone with a spare gpu?

[cursive cape]

nope

Will it be a local AI?

[void shore]

It is meant for local deployment, yes

[cursive cape]

#1372229840131985540 The model created by this dude

[grand raft]

I think the duct tapes has removed, meaning this is a sign that gpt image 2 will released at a moment

[cursive cape]

XD

https://tenor.com/view/shrek-gifgod-smirk-shrek-smirk-shrek-smile-gif-8036133445499881066

https://tenor.com/view/mr-bean-mrbean-bean-mr-bean-holiday-mr-bean-holiday-movie-gif-3228235746377647455

Guys, where are the kimi models?

I think they're removing old models that don't justify their price. Maybe they're going to release an update or something like that? Or they're just cutting back on functionality

oh

[void shore]

second run!

[undone geyser]

Any ai news? Any ai added on direct chat?

[void shore]

once i finish training my new family of models, i am gonna try to get it on a inferece provider and maybe arena will add it

rare chance

[ocean ferry]

Yo u train ur own LLM?

[void shore]

Yea i have

[ocean ferry]

What GPU u used for training that

How much params

MoE or Dense?

[void shore]

A 3090 Ti

and im keeping params a secret so i can suprise people when i launch it

its gonna be way smarter than its params entail

[ocean ferry]

Use RL to make it smarter bro

[void shore]

its a dense

no MOE

[grand raft]

There might be a chance to get Opus 4.7 by Max

[wicked sage]

everytime im bored i just go to arena and ask the ai's to write me a4chan greentext

[wicked sage]

peak as hell, how much did this take to train?

also how much vram do you have

[leaden palm]

the best part of meta ai is the subagents; the worst part is also the subagents (for some reason the ai that's supposed to be naming the subagents ends up trying to answer their prompts)

[void shore]

24gb

[wicked sage]

?! wtf lol

how much ram

[void shore]

A 3090 ti, and 7 minutes to fine tune

[wicked sage]

oh thats awesome

what did you use to train it jic

if i wanna train an ai myself

[void shore]

..?

Like software

[wicked sage]

no no like

hold on i forgot the word

did you use smth like fp4

for training

idk if it was fp4 or something else but Oh Well

[void shore]

Quantization?

[wicked sage]

yes

thats the right word

[void shore]

It was fp16

I think

[thick pawn]

Something need to be done about this recaptcha bs. Can't even use the damn arena because I'm stuck in an infinite loop of trying to verify

[wicked sage]

ah, alr alr

btw how much time did this take in total

like how much time in total did it take to train you the ai

[void shore]

it was a fine tune so about 7 minutes

[wicked sage]

OH

alr alr

[remote dove]

Man I love claude opus 4.7 thinking

Just love how I have to respond 2,000+ times to get claude opus to answer

[wicked sage]

what are you even TRYING to do

[remote dove]

Could be something illegal (if jailbroken properly), who knows 😄

[wicked sage]

🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔 🤔

[remote dove]

Yep you are right

I'm asking how to make a bomb

[remote dove]

But it's for a good reason

So Roblox sucks really bad now

So I thought no one would care if it's just gone

So im planning to make this bomb and hide it very securely and then plant it there and it explodes in Roblox Headquarters

[void shore]

The name of my new ai model is...

Geode Beryl 2

[errant sand]

why 2

[void shore]

The previous Beryl is available on https://huggingface.co/GenueAI

but it is much more stupid

and experimental

[errant sand]

oh

[void shore]

im also planning on Onyx 2

and Thaumite 2

Maybe releasing Geode Pyrite

[errant sand]

whats the difference between all of them?

[void shore]

(yes all mineral names)

[void shore]

mostly model size

[errant sand]

whats the parameters that seperates them?

which one is smarter

[void shore]

Beryl 2 is 1.5B, Onyx is likley 3B, and Thaumite might get bumped up to 13b

[errant sand]

nice

I suggest maybe making the later models use sparse

I heard that around 10% of the model parameters actually do the work while the rest are like I would filler

[wicked sage]

(in minecraft)

[wicked sage]

W name

[void shore]

I know

my orignal model family, was Inelly, but i mostly stole other peoples datasets

so i just moved away from Inelly and now switched to Geode

[wicked sage]

awesome

[errant sand]

what you using to train and build the models? Lambda?

[wicked sage]

also holy crap i just checked

you have a rtx 3090 ti right?? i just wanna be reminded

[void shore]

like GPU rentals?

[wicked sage]

mf thats 2K DOLLARS

🙏

[void shore]

yea'

[errant sand]

yes

[void shore]

BUT

its rented

[wicked sage]

OH

ok that makes sense

👍

[void shore]

since i use Vast.ai, i get it for 14 cents an hour

[errant sand]

oh nice

how many GPU's do you use

[void shore]

One at most

but the gpus i use switch often

like a 3060

[wicked sage]

ah, so if im not wrong thats 1.2k dollars per year

[void shore]

or a 5090 some times

[void shore]

yeah

[errant sand]

is it a H200 or T4 or what gpu

[wicked sage]

actually wait i care about monthly

ill calculate that alsi

[void shore]

my time with them are like 4 hours at most

[errant sand]

I feel like im asking too much questions lol

[wicked sage]

100 dollars per month

[errant sand]

not bad

[void shore]

i like answering questions

[wicked sage]

honestly for a rtx 3090 ti thats not bad

genuinel

[void shore]

like a 3090, 3060, sometimes data center gpus, like A100s

or A10s

[errant sand]

nice

[void shore]

im just waiting for it to train

[errant sand]

I was thinking of using lambda to rent some gpu's for NVIDIA H100 SXM 8x costing around 3.99 an hour

[void shore]

this one is gonna take twice as long so im here

[void shore]

thats a lot of gpus

[errant sand]

thought its not a bad price

yeah thought it would speed up the progress

[void shore]

i dont go anywhere past like 40 cents an hour unless i absolutely need it

[errant sand]

oh wow ok

[wicked sage]

yall over here renting these good gpus while my dumbass is training an ai with a free t4 gpu 🙏

[remote dove]

I train ai through html

[wicked sage]

how does THAT work

[remote dove]

Well you seen the Channel called Ai warehouse?

[wicked sage]

❌

[remote dove]

Search that up and you'll know what I mean

(On youtube)

[void shore]

thats not free for me

i dont get that luxury

[wicked sage]

OH THIS THING

😭

[void shore]

i have to pay

[wicked sage]

i love this channel

[wicked sage]

damn

[remote dove]

That's how I train it

[wicked sage]

i said free cuz im using googlecolab

[remote dove]

I'll brew it

[wicked sage]

i think its free but it not because i have like

88 computed units or whatever

and its going down by 1 each ower

[remote dove]

Introducing Arena.OfficialAi

[wicked sage]

hour

[void shore]

i spent all 200 of my compute units on a really long H100 run

worth it

[wicked sage]

yo im just saying crap 🙏

[void shore]

i trained InellyOSS off it it

my first model

ever

[remote dove]

How does it work

[void shore]

https://huggingface.co/GenueAI/InellyOSS-2.7B

[wicked sage]

inellyoss sounds like a operating system im sorry

[remote dove]

I don't really know how to use huggingface.co

It's very confusing to me

[wicked sage]

i know how to, but setting the thing up sucks

i think

[remote dove]

Can you tell me

[wicked sage]

wait ill give you a tut on how to set the thing up and

use the ais

[remote dove]

Ok

uses the ai

[void shore]

i need a cheap rental to come across so i can make a Genue api

[wicked sage]

i cant find the damn tutorial

[void shore]

to host these models

[remote dove]

Can you make the app

[void shore]

i can make apps

just cant host the models on my own hardware

[remote dove]

Oh I see

[void shore]

unless i buy like P100s

that are really cheap

[wicked sage]

https://huggingface.co/docs/transformers/quicktour

[wicked sage]

wait how does the family form

P100 > A100 > H100?

wait sorry i meant is p100 before a100 and is a100 before h100

[void shore]

P100 is by tesla

tesla gpu

[wicked sage]

ah

ok

[void shore]

its a old 2016 processor under the Pascal archetecture

A100 is a newer one with Ampere (the same in 5090s) released around 2022

and the H100 is more recent, and under the Hopper archetecture

allowing for up to 100GB of Vram

[errant sand]

you can get from google colab for free

[wicked sage]

yeah thats literally what i said LOL

[void shore]

Vast is perm

[errant sand]

they give you one free gpu t4 and one free TPU v5

[void shore]

lkike i can rent my gpu for as long as i want

thats what i like about it

[wicked sage]

wait they give you tpu v5 also??

[errant sand]

how good is it, like would you recommand it

[errant sand]

yes

[void shore]

how good is vast?

[errant sand]

yes

[void shore]

i think its the next step up from Google Colab

similar UI

preloaded credits (like compute units)

and the same notebook interface

OR

you can connect it to your own terminal at home

[errant sand]

like a more better site with more gpus?

[void shore]

but it requires extra setup

[void shore]

a LOT more

like 16000 GPUS are available for rental on average

thats why you can get really good deals

i think sixteen thousand

[errant sand]

and the cost is around .43 cents an hour for one?

[void shore]

the sweet spot is 20-35 cents an hour

because you can get a lot of vram from that

[wicked sage]

ye ye ik you just sent this like 2 minutes afgo

[errant sand]

how many vram like 60?

I thought you didn't see it lol

[void shore]

it depends on what GPU is available

try to go for the consumer gpus

30xx, 40xx, 50xx

20xx, and 10xx are too old

and GTX has like 3GB Vram on average

Go for T4 for datacenters

or an A10

[errant sand]

I'll check them out, thanks!

[void shore]

ill make a gpu guide one day

[errant sand]

share it with us when you do

[wicked sage]

https://tenor.com/view/tyuyos-ishowspeed-ishowspeed-meme-speed-speed-meme-gif-17616160487069326613

[errant sand]

gpt 5.4 high won????

gemini cooks gpt by a large margin

[undone saffron]

The vote says otherwise

[sharp mirage]

yoo hi all

[viscid cloak]

yo friends, still no clue which model flow−state is?

[ocean ferry]

Hi

https://x.com/i/status/2045680175036113059

can (@marmaduke091)

Yep. GPT-5.5 has been in testing on ChatGPT for a few days now, and I can confirm it’s better than Opus 4.7.

X · Apr 19, 2026, 1:45 AM

[wary robin]

yo lowk claude opus and gpt 5.4 high should be in side by side and direct's "max" model router

[polar horizon]

mfs be lying for fun

[wary robin]

no it has

a newer model apparently

I've gotten it as well

[polar horizon]

on everyone's soul gpt 5.5 ain't even better than opus 4.5

[polar horizon]

where

[wary robin]

it asks but doesn't specify that it is gpt-5.5 or whatever. It just said "newer model"

[polar horizon]

https://tenor.com/view/speed-ishowspeed-reaction-reactions-meme-gif-8141100727479891642

[wary robin]

it asks you to compare repsonses occasionally with a newer model

[wary robin]

also 5.4 xhigh is better than 4.5, but not 4.6 or 4.7

[polar horizon]

on arena.ai right

[wary robin]

no on chatgpt

[polar horizon]

ahh

i forgot about ts

been addicted to arena.ai

shii imma check ts out

[wary robin]

mhm

[wary robin]

I only got it once tho

[polar horizon]

so you got it once and you agree it's better than 4.7 opus ?

[wary robin]

no i don't think its better

it was a pretty coding intensive request

[polar horizon]

oh you meant it actually exists

[wary robin]

yeah

[wary robin]

and its in testing

[polar horizon]

ah bett

[wary robin]

for us normal people

[polar horizon]

"normal people"

ts cooking me

[wary robin]

yeah i doubt ur some sort of albert einstein with insider knowledge or ceo brains

[polar horizon]

https://tenor.com/view/tesla-and-einstein-luma-ai-gif-14475294631445060178

id say i got some specialness in me

[wary robin]

also have the credit things been implimented yet?

[polar horizon]

fuh naw 💔

[wary robin]

thank god

[polar horizon]

why bro

we can't use opus

[wary robin]

its gonna b e way more restrictive

like they gonna cut our usage down by 99% but not say it, and then say it is "how the credit system works"

[polar horizon]

the poe.ai method
website is completely buns now

[polar horizon]

you right

if it eventually gets added

I hope it ain't added to the image models

[void shore]

Training Onyx 2 right now

[pale sonnet]

drake or kanye

[sly cedar]

https://tenor.com/view/diddy-p-diddy-jeffrey-epstein-epstein-p-diddy-to-jeffrey-epstein-gif-9527810481976789768

[pale sonnet]

https://tenor.com/view/crine-im-crine-son-im-crine-anthony-mackie-tiktok-gif-10273212848663596705

[void shore]

Oh wait

It is

I messed up the gpu types

An A100 is Ampere

[wicked sage]

hi

[void shore]

They are really fast

[wicked sage]

i am back

[void shore]

Hi

[wicked sage]

🤔

[grand raft]

i AM NOW CREATING LOTS AND LOTS OF PIXEL GAMES USING AI VIA PYXEL

[void shore]

Actually, im changing plans

[wicked sage]

whats up

[void shore]

i dont know what models i should keep or remove

im running out of time on my GPU before i should close it

and im only on onyx

and overwrote Beryl

so...

i might release only Onyx and Thaumite

beryl was never really a focus to me

[errant sand]

what do you mean by you overwrotr Beryl

[void shore]

i saved to the same folder that Beryl was in

for Onyx

[errant sand]

oh

[void shore]

so it overwrote

[errant sand]

are you going to focus more onyx or thaumite more? like whats the difference between the two, is it the same architecture or not?

[void shore]

Its the same arch, yeah

but i feel that Onyx and Thaumite have more possibility in the local AI space

0.5B and 1.5B isnt really that big of a space

but 3B and 7B is

[errant sand]

yeah 3B and 7B is

how often do they hallucinate in their response?

[void shore]

I have added stuff in the dataset to make sure it dosent hallucinate

ill do a hallucination test

[errant sand]

ok

[void shore]

but i dont think it will that much

[errant sand]

try to push them to their breaking point and figure out whats their limit and what to do to make it better

[errant sand]

is it trained on normal datasets like conversional types or is it more a niche specific for example purely math or finance etc?

[void shore]

Its trained off of various logic, cause > effect, math, and coding dataset examples

like a lot

it went from 200 examples to over 1000

[errant sand]

thats good

whats the structure of the datasets you are using?

[void shore]

i cant leak most of it, im planning on keeping it to myself, but it goes like:

{"user": "What is 16 * 19?", "bot": "Multiplying gives 304 since multiply 16 by 19"}

[errant sand]

yeah thats good but probably add some type of indictor of how complex the type of level is for example like a beginner coding question should be something like this

this v

maybe add different difficulty and tags to make it faster for the models to detect the type of questions the user is asking it but this is just a suggestion

[void shore]

i just feel thats more room for hallucination for me

like it tries to type the tags

and its wrong

and it... you know

[polar horizon]

king drizzy all day

[errant sand]

yeah but it should be trained to detect and I thought hallucinations would probably lower if it focuses on specific things/key words like narrowing down topics

[void shore]

yeah

that could work

[errant sand]

how many examples/ entires do you have in the dataset?

[void shore]

about 1000 examples

[errant sand]

nice, like how good are the quality of the examples

quality is better then quantity by a large margin and could severely boost or harm

[slender thistle]

4.7 JB

[void shore]

The quality is decently high for one person

but people told me that i need more data