vibe-coders

[vagrant folio]

yesterday flash was very slow for me for about 4h

the best aproach Ithink is combine all three of them

[stark sapphire]

trying something new. Not sure if i will actually proceed and make it all work.

[stark sapphire]

haha i guess i will make it happen. going all out.

[gentle aspen]

heyyyy welcome @rain lava . Yipeee, you fixed your problem didn't you??🎉🎉

[rain lava]

Yes, I did! .🔥

[rain lava]

Is that an optimizer & security app?

[stark sapphire]

no idea what it is xD i'm just making stuff and see if it's possible.

[rain lava]

What do you mean by "making stuff"? Did you make the app?

[stark sapphire]

I did yes. with stuff i mean i just don't know what i am making yet. Just experimenting.

[rain lava]

Okay

[gentle aspen]

guess who is building a summarizer AI model. it is going to be about ~90m params, it will be opensource and it will be at 8k CTX.

before you complain that you have seing many "summarizer" AI models, this is tiny, and also has a small reasonign chain to filter out only needed part of a text.

[cobalt echo]

💡 Save Gemini API Prototyping! ⭐️
Hey everyone! Google recently (March 2026) restricted the $300 Welcome Credits to Vertex AI only, effectively putting AI Studio behind a paywall for new devs.
We all know AI Studio is way faster for prototyping Gemini 3.1 Pro and handling 4K images (no GCS overhead like in Vertex). Forcing us into Vertex for initial tests just kills the "vibe coding" speed.
I’ve opened a feature request on the Google Issue Tracker to bring the credits back to AI Studio. We need Stars to get Google's attention!
Vote here (Click the Star ⭐️): https://issuetracker.google.com/issues/504375652

[rain lava]

Don't all summariser AIs use a reasoning chain to filter out text?
I mean, that's literally the point in summarisation

[gentle aspen]

no, they dont.
how things usualy go is:

[large corpus] --> [summarized]

although, this is 'fine", it removes the core factor of text.

my AI model does:
[large corpus] --> [reasoning: ChunkX, ChunkX...] --> [final answer]

[rain lava]

What makes you think everyone else agrees though? I know alot of people that like Vertex over Gemini.

[rain lava]

Are u using a bpe or byte tokenizer?

[gentle aspen]

it wont be a crazt architecture, i just want ot see it work, then I am adding formating support. for example, people can get responses in: plain text, MD, HTML, etc

[gentle aspen]

bpe

[rain lava]

But why

[stark sapphire]

so many AI things... i don't even know what Vertex is, i'm googling it, and i'm still a bit confused lol.

[rain lava]

Using a BPE at that small of params is going to tax you.
a 50,000-word vocabulary with a hidden dimension of 512, the embedding matrix alone takes up roughly 25.6 million parameters.

[gentle aspen]

bcs it save memory compared to byte

[gentle aspen]

well, gpt4o sugested me that years ago, and i sticked to it. so i was doing it wrong the whoel time?

[rain lava]

Yeah kinda.
BPE Tokenizers tax the fck out of small models, but is good for training speedx

[cobalt echo]

But Vertex AI is absolutely the right product when scaling production apps and implementing MLOps pipelines. The problem is that this credit can only be used for creating proofs of concepts.

The latency for uploading 4K images through IAM and GCS buckets to train models using Vertex is very large, around 30-60 seconds per image. On the other hand, this process is much faster in AI Studio (instantly). Also, on-demand Vertex endpoints often reach 429 Resource Exhausted errors due to sharing one enterprise capacity across the whole company, which makes "vibe coding" in AI Studio significantly more convenient.
We are not against Vertex AI – but we just want to have the freedom to use our intro credits where it is more convenient to us. If you appreciate flexibility, please give a Star to our issue!
See support thread: https://issuetracker.google.com/issues/504375652

[gentle aspen]

yeah, I will use byte, then.😄

[rain lava]

U might wanna research first

Byte has cons too

[gentle aspen]

yeah, ofc!
I will see.
I should check into deepseed too

[rain lava]

I'm 97% sure u can fit that model in ur vram very easy with optimisations.

Deepspeed is more for multicard setups

It creates vram buckets and it will slow u down

[gentle aspen]

I will try.
bcs i did traina 90m model before at 4k, it used up all my VRAM (without spilling), since I am doing this at 8k (bcs I have to) I might have a hard time. but I shoudl try

[rain lava]

After lookin' at pros & cons, my reccomondation is take bpe but lower the vocab so from 50 k to 16k, this will be less params in matrix

[rain lava]

Did u use deepspeed last time..

[gentle aspen]

hmm, I was going to do 24k, will 16k be enough?

[gentle aspen]

no

[rain lava]

Covers all common english words, wont work on very rare ones

[gentle aspen]

there will almost no rare generation tokens tho

[fiery lagoon]

https://tenor.com/view/skibidi-toilet-spiderman-skibidi-gif-11268458470528183028

[rain lava]

24k Vocabulary: 18.4m parameters.
​16k Vocabulary: 12.3m params

[fiery lagoon]

hmmm

such a good friend

[gentle aspen]

I was gonna do 25m params for the test run

[gentle aspen]

like dont say that shi😂

[fiery lagoon]

nah nah okay i get it....

ur too cool for me now

gold on google

[gentle aspen]

btw @fiery lagoon you might wanan delte that gif

[fiery lagoon]

y

[gentle aspen]

a mod might ban or smthn, idk

[fiery lagoon]

😭 its litteraly just a gif

[gentle aspen]

https://tenor.com/view/son-im-crine-son-im-crine-gif-5176045023480767429

[rain lava]

|Param, Cost, usage of 25m, brain params left|

50k 25.6M 102% -0.6M (Impossible)
24k 12.3M 49% 12.7M
8k 4.1M 16% 20.9M
4k 2.0M 8% 23.0M

[fiery lagoon]

some one should ban you for this

[gentle aspen]

anyways, back to 25m params at 8k. why does CTX always scale O(n^2)? so annoying...

[fiery lagoon]

when u gonna update oncard for linux bro

[rain lava]

When i get into a healthy sleep scheudle

[rain lava]

Idk bro ask geminj 😭 🙏

[fiery lagoon]

https://tenor.com/view/get-a-load-of-this-pointing-pointing-finger-car-wheel-gif-2572264087411888860

[gentle aspen]

GEMINI?

u use gemini?

[fiery lagoon]

gemini is peekkkk

deepseek is peeeek

[rain lava]

This is news to u...?

[fiery lagoon]

chatgpt is asss

[gentle aspen]

no offense to gemma, but full on offense for gemini. isn;t good for what it says it is.

[rain lava]

Ive lit told u i have a geminj pro sub?

[gentle aspen]

agree. I loved gpt2,3,4o

but now..

[gentle aspen]

ohh, then it is totlay fine

[fiery lagoon]

2 3 4 all ass deepseek is on top

[gentle aspen]

i hope you know what you are talking about

[fiery lagoon]

bro im not like u i dont spend all my time on ai

deepseek is just the best

tho

[gentle aspen]

the reason why deepsek exist in the first place is bcs of GPT2

[rain lava]

I definetky find the benchmarks better than the model but i feel google js throttled geminis thinking time, and benchmarks are unthrottled.
Its also not THAT bad

[rain lava]

I also get 5tb cloud, lyria 3, nano banana 2, pro, imagen 4, veo 3...

[gentle aspen]

i mean, i can run gemma4:26b at 256k at usable speeds. and also pared with wikipedia and bing or ollama cloud it is basically all the so called "paid performance"

[gentle aspen]

ik, but I already have a ChatGPT sub, but I am plannign to move to claude

anyways i gtg, i have to download osme datasets

[rain lava]

I can see why tbh, ima look at hpw much ts is

[gentle aspen]

sam price, google is mor eexpensive

my days, I need to get better fingers or a better brain for typing

[rain lava]

17 for claude pro, 33 for google pro, but i see no video generation, or img, or sound, or cloud stoeage

[gentle aspen]

that is with the annual plan

[rain lava]

"$17
Per month with annual subscription discount ($200 billed up front). $20 if billed monthly."

Pro doesnt even get high traffic priority.

Pro shouod be plus, its kinda eh tbh

[gentle aspen]

only thing I hate baout lcaude pro from teh reviews are the rate limits

[rain lava]

Nah theyre all bad

[rain lava]

I was like excited for 1ms cuz i thought max wqs yearky then i saw ots monthly

[gentle aspen]

they are indeed bad, but they are the reason why good models exists in the first place. i just love the innovation they brought with those models

[rain lava]

How is that the reason.

[gentle aspen]

personal bias to OpenAi yk.
I mean afterall they are the reason why transformers are popular besides Bert (which is nott even a pure decoder transformer)

[rain lava]

Sam Altman Glazathon

[gentle aspen]

thats rude.

[rain lava]

You say that when I've done it twice but when other people spam it it's fine? 😭

[gentle aspen]

not liek that, but I dont liek when ppl start being sarcastic with me and my personal preferences. i just assuumes, they respect my choices as mcuh as i do with them

dw, it's just my nature lol :)

[delicate blaze]

starred and upvoted!

[delicate blaze]

id like to talk about vertex.. have a specific project i would like to launch a team of agents into.... i am also very comfortable using Claude Code & Antigravity for almost all of my recent projects. but the costs associated with vertex are not yet totally clear to me and thus i hesitate to get into it without clarity. would you be willing to jump in a DM or invite me to your own server here on discord so we can jam?

[cobalt echo]

You can add me as a friend and send me a private message with your specific question

[gentle aspen]

aight, Ijust made an architectureusing GPT5.4:

Tokenizer:

• Custom byte-level BPE
• Vocab size: 16,384
• No UNK dependence
• Fast Rust/tiktoken-style runtime implementation

Model:

• Decoder-only transformer
• 15 layers
• d_model = 384
• 6 attention heads
• 2 KV heads (GQA)
• head_dim = 64
• SwiGLU FFN, hidden = 1024
• RMSNorm, pre-norm
• RoPE positional encoding
• Tied embeddings
• Full causal attention
• Target context = 8192
• Total params ≈ 29.9M

I am sus abou tthe 15 layers. that looks too much for a 29.9m model, i might wanna recalculate it, but does it look solid?

[rain lava]

Yea looks good, just make sure it doesnt relapse training data or gets a loss to low otherwise you'll get overfitting.

[gentle aspen]

I made sure that there is about 25x the params

for SFT i am aiming for about 100m tokens and for PT, I already collected 500m tokens. i am doing some extended PT on reasonign tokens, so i can support -100 masing in the seciond SFT run without ruining the gradients too much

I just want to play it safe for now

for a little experiemnt. I mean, this is very simple, so it should work

[vagrant folio]

Hi Ag power user in pro, just to confirm Ag also reduce a lot tha flash model usage?

Before I was able use it for 3h withut stop now in just 1 and half i run out of model usage

[ionic mica]

Para que es este server @dapper fractal @uneven echo

[gentle aspen]

can someone help me with this😭

[rain lava]

Contact github support

[gentle aspen]

i have to log into github to do that

[rain lava]

Then do accouht recovery

Also u can contact support, it tells u to even if u cant.

And check ur email fpr anything tos related

[rain lava]

So, how's training going?

[gentle aspen]

still making some reasonign datasets and curating the PT dataset. but I will start training today

but i am more onverned about my github

i am so messed up dude. my github = my life

I NEED IT

pls helpme

[rain lava]

https://docs.github.com/en/site-policy/other-site-policies/github-account-recovery-policy

[gentle aspen]

lets see how things go

might overfit it a bit.. I didnt have enough tokens too, plus it is at 2 epoches

[gentle aspen]

that didnt work too. they dont have any section for this.

all the links lead tot his, no support contant. nowhere... in a site from a billion dollar coporation...

I got suspended for absolutely no reaoson. I DID NOTHING.

[rain lava]

Did u try

https://support.github.com/contact-next

[gentle aspen]

it takes me to this page where I have to put in my number. it says "incorect mobile phone number" EVERYTIME. :(

[rain lava]

Are u in the US though.

[gentle aspen]

no, but i filled that correctly

[rain lava]

Did you try removing/adding a 0 at the start of the number?

[gentle aspen]

yeah I tried all. look what it says

[rain lava]

Ur rate limited.

[gentle aspen]

from what exactly???

[rain lava]

Also try raw numbers, 0 dashes etc.

Ur rate limited from too many ohine num reqs.

[gentle aspen]

I should try tmrw ig

brooo, i cant live without github

:(

😭😭😭

[rain lava]

Rate limits dont take 24h to go away?

[gentle aspen]

github is my social media platform 😭. it all my fun in life brooooooo

[gentle aspen]

really?

[rain lava]

Yes?
Have u literally never got a rate limit on discord? It takes like 10mins

[gentle aspen]

no

never even heard bout it. DISCORD HAS RATE LIMITS??!!

[rain lava]

Heaps

[gentle aspen]

LETS GOOOO!
I got it working, somehow....

it is kinda dumb, i trained it with like 4m ish data for a 270m model, it is absolutely not enough, so i will train more, but to confirm, I knwo this works. NICE!

anyone got tips on improving and making the CoT stable?

[gentle aspen]

I am going to have a fun ~time~

[rain lava]

You're getting the data, normally right?

[stark sapphire]

<@&1009526435276394496> spams in more channels.

[gentle aspen]

YAYY!!!
I distilled claude opus4.6 into gemma3:270m:

I should try it witha harder prompt. btw this is still a test run.

I just wanted to nump up the steps and training data, it was trained on 150m tokens (not a lot)

but the reasonign chain is wayyy stable than qwen tbh.

i will focus on the reasoniign structure, bcs even tho it is mostly opus, i mixzed in soem deepseek data too, bcs i am broke, and opus is expensive

[winter viper]

sup folks, news on opus 4.7?

[gentle aspen]

ehhh, google is doing the hard to get for free users (maybe for paid user's too)

AI in their first few steps are so dumb lmao.

btw guys, i have a question.
is gemma4 using mamba?

[rain lava]

Compare the same prompt to undistilled

[undone pendant]

Still no dev update on Opus 4.7? Are they working on it?

[undone pendant]

restart for update 😮

1.23.2
Apr 16, 2026

Bug Fixes
Fixed bug that prevented MCP servers from loading and bug that prevented accessing workspace-specific settings.

Improvements (0)
Fixes (2)
Fixed bug that prevented MCP servers from loading
Fixed bug that prevented accessing workspace-specific settings
Patches (0)

--

Huge

[rain lava]

Dude stop winging, your bio says "full stack developer". If you feel the need to use that in your bio you also should be able to code yourself without the use of Artificial Inteligence🙄

[rain lava]

Do I look like a google employee to you? Why would I know when they're working on it.

[gentle aspen]

fr.
bros stack:
claude code: backend(server)
codex: frontend(client)
cursor: backend (client)
opencode: frontend (server)

new bio:
I am a <buzz word><buzz word><buzz word>. I like to do <buzz word> with AI. bcs I dont know how to do <buzz word>, so i use Ai to do <buzz word>.

[rain lava]

Fr lmao 😂

But you forgot to add to his bio
"But i can do scratch" 😭

[gentle aspen]

making some progress on my 30m model (stilld oing PT)

should I opensource the weights?
it is:

SP BPE 16k
30m @ 8k
casual LM
decoder only transformer, nothing crazy.

[rain lava]

Do what u want.

[rain lava]

THIS is scratch btw

[gentle aspen]

...But I can do scratch

watchign YT tutorials...

no no no, this is only for education resources...

[gentle aspen]

man, i hate scratch. I mean, no ffense, it does help you improve engineering reasonign, but it makes you think developing is all sunshines and rainbows, and makes yu not appreceate actual programmars.

i mean, I shouldn't be speaking, bcs i use AI, but respect to all the people who actually put some work.

[rain lava]

😭,

[rain lava]

I dislike scratch to, alot tbh, its so annoying,

[gentle aspen]

fr.
teachers glaze it too much like it will get you a bachelors degree

It's not a 5 marks question bro😭

[rain lava]

"Everyone! If you don't learn scratch you will NEVER know anything. Scratch is the key to making any games u want!!
Scratch is very hard, i know, but if u guys can do it you'll all be unstopable! ⭐"

They also never know what programming & scripting languages is. Mention "python" once and their like "whats that!"

[rain lava]

Me when hallucination

What in training its also asking questions..?

Also whats the loss rn

And did u give it enough data to not overfit yet

[gentle aspen]

ohh, and also:

"Kids, if and elif is everything you need. helloworld("print") will get y a long way. you are the next sam altman, sundar puchai, elon musk" 🥀

[gentle aspen]

its just an eval, so check the ppl

[gentle aspen]

about ~2.7

[gentle aspen]

I followed the chinchilla scaling law: 20 x Parameters = tokens neded

I did it in a scale of ~25 x params

istg, if someone asks about sum "opus4.7 on AG" I am fr going to crash out bro

[rain lava]

Its like that one python programmer who onky knows python and will always say "most people know if, and else, but not many know elif!"
And proceed to say python is js the best panguage.
Also "helloworld("print")" ?! 😭

[rain lava]

Them not knowing u can use claude code with google vertex ai, which is apart of the pro sub.

[gentle aspen]

"Most will say if and elif is what they know, but I am not most people..." ahhh🥀😭

[gentle aspen]

I mean. what are they exp[ecting bro?

to give you the best coding model for $0?

[rain lava]

Exactly 😭

Complaining without knowing how good they have it

Chat can this card train llms 😭

[gentle aspen]

what is it?

[rain lava]

RX 580 Strix

[gentle aspen]

🥀🥀

[rain lava]

Paired with a

[gentle aspen]

I feel bad for you...

out of all the cards you can get, you got that?

[rain lava]

Thats not whats in my pc theyte just spare parts 😭

[gentle aspen]

okay, that changes the whoel story

[rain lava]

Whats wrong with an rx 580 ?

[gentle aspen]

it's old and unsupported by almost all the ML libraries and, it can barely do anything

[rain lava]

As if i coukd even attempt to train a 1b model on an rx 580

[gentle aspen]

it's 6GB right?

[rain lava]

8gb

[gentle aspen]

you cant train a 100m model on that at 2k

[rain lava]

100m on an rx580?

[gentle aspen]

will take you more than a year if you tried to train a 100m model

[gentle aspen]

abslute bare. will take you months

omds my typos

[rain lava]

I think this is conparable to an 1060 or 1070

So it can play games atleast

[rain lava]

I have a spare 1070ti too

[gentle aspen]

not at all!

nvidia has their own pipeline, and software stack for training NLP models (or anything in general tbh)

a similar GTX will be wayyyy faster

[rain lava]

Or non ti idk

[rain lava]

I meant in games

[gentle aspen]

okay NOW we are talking.

that is good.
nothing crazy, but "okay" for some light trainng

[rain lava]

Okay but that wont compare to my 9070xt, right?

[gentle aspen]

not ba, but dont expect to crank up the settings in FH5.

I mean, you already have a 9070, right?
I remember you said something liek that.
that is WAYYYYYYYYYYY beter

[gentle aspen]

no shot sherlock

[gentle aspen]

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

[lime sentinel]

Not even close 😄

A 9070 XT is in a completely different league, way newer architecture, massively higher compute, bandwidth, and proper support for modern ML stacks.

[rain lava]

My 9070xt can run fh5 on max rt, extreme graphicsk and 8x msaa at 120fps 1440p

[rain lava]

Ik i was making sure OM didnt think that lol

[gentle aspen]

https://tenor.com/view/sherlock-holmes-robert-downey-jr-firm-grasp-of-the-obvious-obviously-obvi-gif-3604034

[lime sentinel]

A 1070 Ti is okay for light experimentation, but the RX 580 is basically legacy for ML at this point.

[gentle aspen]

ikr

[lime sentinel]

For gaming, yeah RX 580 ≈ 1060/1070 range.
For ML, NVIDIA still wins hard because of CUDA + ecosystem.

[rain lava]

Ther3s a chance the 1070ti is broken tho!

Its 5 + years old used

After my father upgraded his computer i gutted the old 1 for that gpu

[lime sentinel]

Vintage hardware build: comes with nostalgia, uncertainty, and random crashes included.

🤣

Either it boots… or it becomes modern art real quick

[rain lava]

I was using the old pc and then it js crashed, re installed windows and holy, the boot screen turned orange and blue n shi

[lime sentinel]

Orange/blue artifacts on boot? Yeah, that’s less Windows issue and more GPU VRAM waving goodbye😂

[gentle aspen]

why do you sound like chatgpt😭

[signal raft]

ChatGPT 4o ahh, but it is goated

[lime sentinel]

If this is ChatGPT, I need a refund latency’s too high.

[rain lava]

1minute replies bc the server traffic is full

[lime sentinel]

before AI humans were good and they are good not better than AI but not even worst than AI

[signal raft]

That issue is Antigravity only xD

[lime sentinel]

Yeah, free tier struggles 😔

[gentle aspen]

"You ran out of Human2026:rohit model. Your quota resets next week"

[rain lava]

I don't get that issue

[signal raft]

I mean it is obvious, they want you to use the cheaper Gemini Models

Instead of Opus or Sonnet

[gentle aspen]

for me, codex performs really wellcompared to AG.

AG is too confusing and delivers less for what it is.

[lime sentinel]

Upgrade required: sleep + caffeine pack.

[gentle aspen]

yeah, but AG doesnt let me use gemini flash tho

Sam Altman lets me use anything

lol

[rain lava]

I feel claude models are js the best but theyre expensive to run so less prompts

[signal raft]

Yeah I think so too. I have got myself the 100 Euro Plan for Codex, even though I didn't feel good giving Sam my money xD

[rain lava]

Js use gca then

[gentle aspen]

yeah. Plus gemini models are KNOWN for a bad CoT.

gemma has a better CoT than gemini

[lime sentinel]

Same experience here. OpenAI Codex tends to be more direct because it’s optimized for actually executing and iterating on code tasks end-to-end, not just generating answers.

AG feels more flexible, but that often comes at the cost of clarity and signal-to-noise. Codex just gets to the point and ships usable output faster.

[rain lava]

How google felt putting a non coding-at-all model into an ide

.

[gentle aspen]

https://tenor.com/view/patrick-drooling-patrick-star-spongebob-movie-drooling-slime-gif-525340425450564005

also like this👆

[lime sentinel]

Feels like a UI decision that got ahead of the model capabilities.

[rain lava]

Ppl were arguing abt when i said google wasnt an ai first company, which i was wrong but thats the irony 😭 a company making sm ai then making shi ai

[gentle aspen]

fr. they wanted it so it is like:

"So, we know it can't code well. we, uhh we uhu h huhhh [hallucinaton]"

[rain lava]

Maxamised cheapness over quality is what they did

Limited its thinking time i thijk tbh, and unlimited it on benches

[lime sentinel]

More of a product problem than a model problem they shipped the integration before it was actually useful.

[gentle aspen]

it's crazy how google cannot train a good AI model, they have ownership to many sources of information and also own large datasets.

they are indie devs who makes better models.

at least finetune a coding first model... not that hard considering that itis a billion dollar company.
I hate it when Ai companies think their singlemodel can do EVERYTHING.

[rain lava]

Trillion*

[gentle aspen]

https://tenor.com/view/elmo-sigh-gif-5895699729383201961

[lime sentinel]

That’s a bit harsh tbh. It’s not that they can’t train good models they clearly can. The harder part is aligning one model across multiple use cases search, docs, code, without degrading performance in any one domain.

Indie models often feel better because they’re narrowly optimized coding-only, while big companies are trying to balance generality, safety, latency, and scale at the same time.

[gentle aspen]

dude, OpenAI started with nothing. I am not glazingopenAi, bt if they came to this point, google has all the reasons to make a coding first models.

also gemini isn't bad by any mean, but they dont properly integrate it.

if it is free, at least they should warn about an estimation of when or how my tokens are going to run out. they just say "quota bye-bye👋" and we have to guess...

[lime sentinel]

Agree on one thing though forcing a single model to do everything usually leads to mediocre results in specialized tasks like coding

[gentle aspen]

THATS WHAT I AM TRYNA SAY!

[lime sentinel]

Yeah, turns out it’s not about having resources, it’s about focus. Coding-first models work because they actually… focus.

[rain lava]

Apparently qwen3.6 is very good at anything, i dont even see a qwen3.6-coder.

[gentle aspen]

yeah, that is true. i testedit out. they didnt mean "best of all" they meant "best, considering it';s size and the fact that it is opensource"

[lime sentinel]

Good at anything usually just means not specialized at anything tbh.

[gentle aspen]

also people can finetune it easily unlike gemma4.

the audio embeddings takes too much VRAM

[gentle aspen]

actually, it is really good

[gentle aspen]

you should try it

[rain lava]

Yes, i feel 3.1pros arch is speed optimised, which is why 4.6thinking was better with that 3phase reflection process that thinks

[gentle aspen]

and the plus model is free at their websiote. dont know how they did that. tbh china has better infrastructure when it comes ot making things efficient

[rain lava]

Its not anywhere near frontier models, im pretty sure its 358b

[gentle aspen]

I mean, the 3.1 is "okay" but it is more of a general model.

[lime sentinel]

I’ll benchmark it myself really good tends to depend on the workload

[rain lava]

A general model marketed qs coding math etc

[gentle aspen]

35b-a3b model is the opensource one. usually their big models are always above 200b

[rain lava]

Im talking abt 3.6plus. U said its free, i say its cuz its not 1t params, its 358b

[gentle aspen]

it is a genera model with more coding and math data, that doesn't make it "best", but definietely better than many others, but still struggles even with a long CoT

[gentle aspen]

358b is not enough for you? 🙄

[rain lava]

When did i ever say that dude

[gentle aspen]

after about 20b, it is all about data, and not about parameters.

an older model (2024 ish) can do really well with good coding and math data

[gentle aspen]

here

a model does not have to be 1t to perform good yk

[rain lava]

8m just proving that its free due to a lower param count there? Where did i ever say 358b is bad?

[gentle aspen]

ohh

[rain lava]

I wanted to see that 😭

[gentle aspen]

here is a better view at that beauty

[rain lava]

What makes a dataset rlly good?

[gentle aspen]

structure

[rain lava]

Define structure

[gentle aspen]

instead of:
let me think this through. I shoudl summarize this. some points would be to talk about how neural networks are ood at understanding semantic relations ships, and maybe even how ML works would work!

compared to:

let me think this through.
We got:

Neurl networks

I can talk about:

• how they understands semantic relationships.
• How machine learngn works
• modern ways of re-inforcement learning.

[rain lava]

Ah okay

[gentle aspen]

what in epstein is this bro???😭

[gentle aspen]

I have ebing looking at this the whole day:

bro really said "ebing"🥀

[rain lava]

If my google nest-mini says "Sorry, I don't understand" when I ask a question, am I smarter than it?

[gentle aspen]

yayy, i get 2 more TPS when i put my RAM to 4800 --> 6000

it was runing parts of the tensors in ram. ykw? I shoudl run qwen3.6 and see

[rain lava]

U can also turn on high bandwidth mode.

[rain lava]

Nvm it doesnt increase bandwidth, though it does optimise timings but is unstable, its a silicon lottery thing, itll likely work on sk hynix ram chips

[rain lava]

Or maybe it does, idk, apparently it can be better for running/training, but data loss will occur if its unstable

[gentle aspen]

I mean, my pc is already capable, so I will choose stability oover speed. righ tnow, everything feels smoothf or me, so i have no oproblems.

also I should install Linux, man.

windows is gettings on my nerves.
WDYM I CANT USE FLASHATTENTION😭

[rain lava]

If ur ram uses sk hynix and u run mem test with the high band mode u can get 10% or smth increase

[gentle aspen]

dang...

[rain lava]

My parents use a 6400 cl32 2x32gb kit for webbrowsing whilst im stuck on 5200 cl40 ... I could get a 35-40% increase if i get their ram

5200 CL40 (Stock, XMP): ~72 GB/s
​6400 CL32 (Stock XMP): ~88–90 GB/s (+25% increase)
​6400 CL32 + HBM Enabled: ~98–102 GB/s (+38% to +41% total increase)

[gentle aspen]

FOR WEB BROWSING????
what kinda web browsing requires that?

4800 mt/s is more than enough

[rain lava]

My dad saw a salefor office pcs and saw 64gb ram and clicked buy

"Office pcs" is a stretch, theyre more workload capable.
Rx 6600XT Eagles.
I7 14700kf
32gbx2 superram

[rain lava]

Ddr4 can be used tbh

16gb

Okay maybe not 16, my parents love 7million tabs

32

Or 24

[gentle aspen]

yeah ofc they do. tell me a single human who doesnt. i feel like 24gb is the sweet spot. i meant 32gb is fine, ddr4 can handle everything you througha t it perectly. there is nothign wrong with that.

[rain lava]

They use chrome tho, my mums laptop had 16gb ddr5 and she complained it was slow for her 98million CHROME tabs. Like if thwy used non electron and a less sandboxing engine typa browser itd use less

[gentle aspen]

every browserusing a form of browser sandboxing, which means it WILL cost soem ram even if it was not electron. not even webkit can be efficient

[rain lava]

Helium is less expensive than firefox and uses blink still (same as chrome)

[rain lava]

Midori, Falkon, Netserf, Palemoon are all light weight browsers

[rain lava]

800~900MiB for a Gemini Tab on Helium vs 1.0-1.1GiB on FireFox

[rain lava]

To be fair i used google se on ff and duckduckgo on helium to try

[stark sapphire]

Who here uses Cursor over Antigravity?

[signal raft]

I think most people agree that Cursor is the better overall Editor, the only good thing you get with Antigravity is refilling usage limits

[stark sapphire]

yea this, but I think Cursor also gives you more freedom, you can choose out of more models, and it even allows you to connect your own local AI, which Antigravity can not.

[verbal axle]

Yoo

[undone pendant]

Full stack v dev. V for vibe

[gentle aspen]

hey man, i love how you are building something nice!
but this stuff goes under the "marketing" category. yo might get banned. so i recommend you delete this message and share it in general.

[sand vault]

i was thinking to use cursor until i just saw they just struck a deal with SpaceX xD

https://klipy.com/gifs/the-office-walk-away--k01KPTW988PXSTJA2BHFZ01XXR7

[gentle aspen]

Whats wrogn with SpaceX?

[stark sapphire]

see #general lol

[gentle aspen]

yayyyy, 27b Qwen3.6

bruh... 600b, and no vision, or 256k ctx

[rain lava]

I think I'll sell my ram and buy this really good kit.
It's a CUDIMM, 8400 MT/s, CL40 24Gb×2 kit, and it's only like 700~800 bucks.

After I sell my current RAM or prices go down it'll be alot cheaper ofcourse.
I can run at 8400MTs with my CPU if I'm lucky, but if I'm not lucky I'll just run 8000~8200 MTs with some tighter memory tightenings.
It's effective memory bandwidth with HBM (High Bandwidth Mode) should be around ~130Gbps compared to my ~60Gbps.
And because it's CUDIMM it should work on my motherboard at that speed (If not, I'll downclock until it works with tighter timings)

State | Speed | Theoretical Max | Effective Bandwidth

Current (Corsair 5200 CL40 UDIMM): 5200 MT/s 83.2 GB/s ~62 GB/s
New (8400 CL40 CUDIMM): 8400 MT/s 134.4 GB/s ~118 GB/s
8400 + HBM: 8400 MT/s 134.4 GB/s ~131 GB/s

[gentle aspen]

48 GIGS ARE 750 BUCKS????!!!!

[rain lava]

Dude it's an state of the art kit

[gentle aspen]

you were able to get like 128 gigs for that price

[rain lava]

No?

Curriencies bro

[gentle aspen]

yes

you did

[rain lava]

AUD vs whatever u use

[gentle aspen]

i thought you meant USD

[rain lava]

No, I'm not american.

Then this one, it's less at other stores but I prefer JB

Metric | Current (5200 CL40) | New (8400 + HBM) | Percentage Gain

Effective Bandwidth ~62 GB/s ~131 GB/s +111%
Inference (CPU/Offload) ~4-6 tokens/s ~8-11 tokens/s +85% to +105%
Training (Offload Step) Baseline -- +35% to +45% Speedup
Inference (VRAM only) Baseline -- +3% to +5% (Negligible)

[hexed marsh]

hello! i have small problem with Antigravity. the latest update made the agent auto apply all changes. before i could review the changes it made nicely line by line, accepting and rejecting them. now the changes get autoapplied and i can only see the comparison that i can comment on, without ability to revert specific lines.

is there anyway to turn this back?

[vagrant folio]

In setting agent command the are a section commands for autoexecution and review, also check reviw policy

[hexed marsh]

i never touched settings, its already on ask for review

[stark sapphire]

Hmm does it never asks to you if it shall proceed?
Sometimes it indeed does proceed, but it only does it so you can work faster for certain small scale things.

sometimes my agent says "Let me proceed with this" without my input. Which is fine with me personally.

[hexed marsh]

no, i dont have issues with agent, with the ui

before it showed the diff directly in the file it worked on

if agent would change this line from True to False for example, then in the editor it would show like this, only with the first one hilighetd in red and second in green, and i could accept it from here

oh nevermind. it works today xD

[stark sapphire]

lol

[hexed marsh]

i guess i hit some limbo yesterday

i'd swear it didn't do that yesterday

or i was drunk and opened different ide

[gentle aspen]

WOW! you messed up soething even a drunk man couldn't.

maybe you just set a new record...😂

[daring ravine]

Hey i am on pro , antigravity, gemini, but it fails. Does it work for you?

[gentle aspen]

btw, does anyone know some good embedding model for general text?

I like nomic embedding bcs of it's opensource license, but I need soemthing small and good like qwen3:0.6b embedding model

[gentle aspen]

https://tenor.com/view/homelander-puff-meme-fat-gif-11878092523812931740

[serene crow]

I did a Proof-of-Concept using Google AI Studio.

TidyUp AI analyzes and transforms disorganized spaces into clean, organized, and structured environments.🧹🌟
#buildwithai

[gentle aspen]

I should make grammarly for vibecoders

it will make your prompts better.
is that a good idea?
also i will allow it to run Ai models locally or let the user choose models from providers like Openrouter, Ollama and someother stuff

[sand loom]

We’re kicking off a brand-new AMA series with the minds behind Gemma 4! 🌍

<@&1043109761299783771> <@&1493624938207379649> Join on #1041705871723466792 our Google DeepMind’s @spring halo & Ian Ballantyne on <t:1777384800:F> for an inside look at how the next generation of AI is moving beyond chat and into real-world action.

Upcoming Sessions:

• 🇪🇸 Spanish: May 12 | 6PM CET w/ @neon oar
• 🇵🇹 Portuguese: May 14 | 5PM UK / 12PM EST w/ @spring halo & @fallow pilot

Don’t miss the first-ever Gemma 4 AMA. See you there!e

[hollow river]

cool

[snow heath]

damn, explanation of a benchmaxxed model

[pearl thunder]

Let's see what we will leave this event ♥️

@sand loom

[sand loom]

We’re kicking off a brand-new AMA series with the minds behind Gemma 4! 🌍

Join on #1041705871723466792 our Google DeepMind’s @spring halo & Ian Ballantyne on <t:1777384800:F> for an inside look at how the next generation of AI is moving beyond chat and into real-world action.

Upcoming Sessions:

• 🇪🇸 Spanish: May 12 | 6PM CET w/ @neon oar
• 🇵🇹 Portuguese: May 14 | 5PM UK / 12PM EST w/ @spring halo & @fallow pilot

Don’t miss the first-ever Gemma 4 AMA. See you there!

[wise rock]

Invalid Invite

[sudden beacon]

Thanks Bruno, we're fixing the event! This is the one meant to be: https://discord.gg/mS7wpPj5?event=1495901248434606080

[wise rock]

Tnks Lorena!

[robust geode]

Hey guys , what's this channel is about, like what I can share in it

[wise rock]

#vibe-coders
This is the place to all your Google Antigravity & Jules discussions!
Download the public Antigravity preview: https://antigravity.google/, and learn more about Jules: https://jules.google/

[hallow carbon]

Hey yo

Actually I got a mail from Google Student Ambassador

I don't know where to ask regarding this

Can someone help me with this Elders

[vagrant grove]

I applied for it today

[hallow carbon]

Mine was applied by Previous GSA

I was busy in work 😢, Working over things

I needed help - So he helped me in filling things

I did the video

And finally uploaded

But I made an issue

😭

Well this is what I got yesterday,

😭

He said they might disqualify me - I am not sure, Please Moderators help me

I faced this error

I shared both images of mail - I received on first day and the one i received on the 2nd day

[stoic citrus]

No one here can help out unfortunately as its a vast community so none knows whose part of what, better contact previous gsa or the program team regarding it and this is the not the channel for this discussion

[hallow carbon]

Well thanks for the clarification Brother

[gentle aspen]

Is there a better stack than electron and JS or python and Qt?

Python and Qt is just a fever dream. man, it HAS to be the buggiest thing y'all (inc. me) has ever seing in our ENTIRE life...

[magic dagger]

Develop a webapp

not everything needs to feel native

[silent night]

hey

does google just reset all quota for all antigravity users ?

[balmy depot]

Looking at my depleted quotas, don't think so. But not sure how you meant that?

[silent night]

idk

my quota just got refreshed all

[balmy depot]

Must be nice.

[silent night]

yesterday gemini 3.1 pro high/low, claude sonnet/opus 4.6 was 3 days remaining

[balmy depot]

I'd show you mine, but it's kinda depressing.

[silent night]

I check 2 another pro account

both of them still exhausted

only this account got refreshed

but somehow I'm experiencing a bug

[fossil estuary]

Hey guys I need jules review

[gentle aspen]

yoooo anyone tried GPT-image-2 or GPT5.5

I think both are beating the anthropic's and gogole's so hard, they might cry😭😭✌️
never seing this powerful models.
tbh, AG should add GPT5.5, it is just... do what you say like opus4.7

[rain lava]

You're saying GPT 5.5 beats Opus 4.7 Thinking?

[gentle aspen]

yes sir, thats rght

[gentle aspen]

absolutely SMOKE it.

[rain lava]

Does it cost money to try and are there benchmarks.

[gentle aspen]

well, codex is now free, and also you get better rte limiting than some software (you know..)

[rain lava]

:0

[gentle aspen]

also GPT5.5 is twice more expensive, but ayy more token efficient, so it is either the same money spent as gpt5.4 OR less, bcs it spends less tiem reasonign

[rain lava]

Does that also beat Mythos Preview?

[gentle aspen]

almost...
You see, mythos is a 15t MoE model, and Anthropic is known for making their models more "creative" instead of Logical. with the creative thinking + better reasoning accross better semantic understand per given context, it performs somewhat better in terms of understanding and iterating.

now, tell me if you (or me) have the money to pay for a 15t model from anthropc while it's Opus4.6 (fastt) model itself costs about 150$/1m tokens😭✌️

[rain lava]

15t is such a waste tbh

[gentle aspen]

the only thing which gemini does better than any other model which is not gemini is... drum rolls... research... thats it... only that :/

[gentle aspen]

well, each model has their own advantages. also 15T is not a dense 15T, it uses sparse MoE with a better internal architecture. dont know what they mean by "interna; architecture" but it looks like it an uyltra efficient 15T

[gentle aspen]

also it is not for normal use. it is to find bugs across very niche stuff which no model will EVER find.

[rain lava]

Well yeah, gemini is developped by google, google is one of the most biggest SE's. And they made a deepresearch model too

[gentle aspen]

No, what i meant was it exlains what could go wrong in what i do instead of saying: "Hey, man, eating raw molten lava is a fantastic thing! The fact yo thought about this puts you ahead millions of people. Short Answer: YES, do it!"

tell me if thatt sounds like something

[gentle aspen]

https://tenor.com/view/war-gif-18347964941877609900

[rain lava]

U can giv chatgpt a custom instruction set to not agree wit u all the time

[rain lava]

What do i get for free

How much prompts with avg token usage

And what model?

[gentle aspen]

everything in codex with slightly better rate limits than AG

[gentle aspen]

dk, i use plus plan

[gentle aspen]

yeah, but still chatgpt's style doesn;t feel natural. it felt more natural with GPT4o and GPT3.5 tbh

[rain lava]

Ok but benchmarls arent everything... (talking abt gemini 3.1pro here)

[gentle aspen]

yeah, i am telling this with user experience, it is actually good.

GPT image 2 is the ONLY model in the world that can geenrate extremely accurate very tiny text.

dude some guy made a very accurate text which is 5-10px in a 4k image 😭

[rain lava]

I think they tried to make it sound more fun or smth with the emoji slam and friendly as hell tone

[rain lava]

Yeah but still no sound model and theyre video model is dying very quick, also no cloudstorage and idk abt opsrc models

[gentle aspen]

nahh man, RLHF wasn't popular back then, so the human chaos wasn;t properly filered with the SFT, now they punish the model for giving very "normal" text, and reward them for giving some bs text :(

[rain lava]

Idk how they raised quotas, they literally cancelled a massive ram order

[gentle aspen]

like i always say, there is no "best platoform" or "best model", each has their own advantage for a certain time period.

as of now, for me the best coding model is GPT5.5 and GPT image 2.0

[rain lava]

U set custom instructions to sound like a tiktok kid bro, did utry non custom...

[rain lava]

What ctx is 5.5?

[gentle aspen]

bro, just dont care abou thte coporate stuff.
if they give soemthing, just dont question it. Who cares if they cancel RAM, buy RAM, OR EVEN MAKE RAM (maybe ppl will care if they do that). either ways bro, just us it.. it's free.

beggars can;t be choosers
if google limits claude usage on AG, I am limited in claude usage. if openAI removed codex, i dont have codex.

simple as that.

[gentle aspen]

in codex default it is 256k, but with fast mode (better intelligence and 1.5x speed at 2x token usage) it is 400k-1m

[rain lava]

You dont wanto question stuff?
Maybe it's deeper than "they gave us free stuff".
Everything is deeper than it seems,

[gentle aspen]

OpenAI is known for 256k-400k ctx.
Gemini is known for 1-2m
Meta is known for 2m
grok is known for 1-2m
claude is known for 64k-1m

[rain lava]

Is 5.5 more expensive than 4.7?
Also i find that 4.7 could be better sometimes, maybe in bigger repos, with more ctx

[gentle aspen]

yeah, if you try to dip into those stuff, tbh you will your time than learning new things, thats why i only focus on things I know that i will get a proper answer.
i mean, researchng about atool is good, but sometimes... just use it

[gentle aspen]

it is WAYYYYYYYYYY cheaper

[gentle aspen]

yeah, it is.
LIKE I SAY EVERYDAY: there is no best model

gemini is good for planning, GPT is good for frontend, claude is good for backend.

sometimes there might be other models which re good at other tasks.
no matter how much they market their models as "good at everything🥀💔" IT IS NOT. period.

[rain lava]

"Best coding model"

[rain lava]

Yay

Why dont companies realise params arent everything? A 15t model, and over t models, are still competing with a qwen 358b model.

[gentle aspen]

BTW, is my new UI better than the old one?

[rain lava]

Which is new which is old...

[gentle aspen]

because after soem point the same data is useless. more params gives the model "creativity" if you would liek to cal lit.

more params means, the model will get more headroom for more context, and more data, also higher params mean it has better semantic understanding. if those benchmarks was something a bit harder. the big models will getmore and the smaller models will get absolutely SMOKED!

[rain lava]

Istg discord is using some ai compression shi, the words are all jumbled

[gentle aspen]

bruh

[rain lava]

U cant tell me ts isnt ai compression

Lnler, suve,

[gentle aspen]

that looks like lossy compresion to me

it is more liek bi-cubic downscaling

[rain lava]

I get suve, whats lnler

[gentle aspen]

thats why the low res inmage has the squar-ish looks instead of a smooth anti-aliased look

[rain lava]

Suve is an a but reducuced sm to an u

[gentle aspen]

I had a stroke reading that

[gentle aspen]

huh?

[rain lava]

Look at my img, theres a word called lnler and suve

[gentle aspen]

maybe a problem with your discord client. bcs when i recieve images it s crystal clear.

did you update hte app?

[rain lava]

I js updated,
Still ahh

Unless they somehow messed up the android port (even tho its fine linux) idk

[gentle aspen]

ohh wait, I have the same issue with my redmi pad SE too

[rain lava]

But whys it hitting 70%, are they even using NEW data?

[rain lava]

Thats android, right?

[gentle aspen]

not shot sherlock

[rain lava]

Ive never used a redmi gng

[gentle aspen]

bcs they are doing the same benchmarks. if they did a different (and by different, i mean HARDER), then the small models are COOKEDDDD

[gentle aspen]

they are "okay", btw it is my first tab, and it is for school, so i dont really care about it

[rain lava]

Yes but with that much knowledge they should ace the EASY benches, no?

Well, moe arch.
Unless theyre trained mostly hard stuff and not much easy?
If so, howre they not fitting easy stuff in 15t?

[gentle aspen]

no.
it is because in real world they wil always fail on those, because even fi they are 15t, they are still train to predict the next token on HUMAN DATA. and humans are MESSY.

unless those companies has the money to clean 15trillion X ~20+ = total words needed, then all the models will get 100% on the benchmarks.

[gentle aspen]

bruh...
let me ask you a question.

tell me what timeline is longer:
[2018-2026] or [back then - 2017] ?

thats why the data is outdated. people fix it in SFT and RLHF.

[rain lava]

Define, back then

[gentle aspen]

<@&1009526435276394496> Hey, how did I get a timeout?
I didn;t do anything wrong.

plus, i was trying to explain something to a person.

[gentle aspen]

like from the 1980s

[rain lava]

Prolly an automated timeout via a bot for a word idk

[vital acorn]

I made a project in which everything is done client side. No actual backend, just pure JS, by choice.

But still want to make it as painful as possible for anyone trying to inspect or snoop through the network tab. No API keys are exposed but there are API calls I'd rather keep private.

My approach is layered annoyance rather than actual security. If you defeat all my layers, honestly respect, consider it a gift from my side.

The layers:

First, if DevTools is undocked and screen dimensions look unnatural, redirect with a "your system appears compromised" message. Casual snoopers give up here.

Second, JS bundle is heavily obfuscated with split Base64 encoded prompt chunks scattered across files. No plain English searchable anywhere in the code.

Third, network tab gets flooded with decoy requests that look legitimate. Finding the real API call becomes a needle in a haystack.

Is it perfect? No. A determined 2% will get through eventually. But 98% of casual people trying to copy the idea will hit the first or second layer and walk away.

[gentle aspen]

you copy pasted that text didnt you?

[vital acorn]

Yes

[rain lava]

Considering there was no "typing" they either did or used a vencord plugin.
Though they just said yes

[rain lava]

1980-2017 is longer,

[gentle aspen]

anyways, ima go get a lil break. i have being vibe coding too much. i should take little break by vibe coding, which i can go back to vibecoding after a small vibe coding break while vibng to some music

[gentle aspen]

EXACTLY. that is more data.

[rain lava]

With newer technology we should be able to make data faster though.

[gentle aspen]

recent data is licensed and hard ot scrape or extract, but the old is opensource, free and unlicensed, so they can easily obtain it. thats why many models are outdated

[gentle aspen]

synthetic data is "okay", but it is expensive, and absolutely not enoughto feed a 15T model bro. plus you need reasoning data too

[rain lava]

Taking a break from vibe coding by vibe coding so you can vibe code after you vibe coding break while vibing to music, not confusing at all.

[vital acorn]

Is this approach reasonable or do you guys know better client side only tricks to make inspect or network tab access more painful

[rain lava]

Ask gemini

[vital acorn]

Already asked Claude, so I don’t think Gemini would know either if Claude had no idea.

[gentle aspen]

tbh, that is some data which you can ask AI, bcs it is not something which requires real world data. plus, an Ai will be more helpful in that matter :)

[gentle aspen]

Gemini is good at research and pointing things out

[rain lava]

Layer 1: The DevTools Undocked / Dimension Check
​The Reality: It’s a great filter for casual snoops. However, a determined user will just keep DevTools docked, open it on a secondary monitor with normal dimensions, or simply use a local proxy like Proxyman, Charles, or Burp Suite. A proxy intercepts the traffic before it even hits the browser UI, rendering DevTools detection completely moot.
​The Verdict: Effective against tourists, useless against anyone who knows what a proxy is.
​Layer 2: Base64 Prompt Chunks & Obfuscation
​The Reality: This stops people from simply Ctrl+Fing your source code for keywords. The fatal flaw here is that no matter how heavily you obfuscate the string, it eventually has to be reassembled and handed to the browser's native fetch or XMLHttpRequest function to make the call.
​The Bypass: An attacker doesn't need to deobfuscate your code; they just need to monkey-patch window.fetch to console.log the final, clean URL and payload right before it gets sent out.

[gentle aspen]

Claude is better at more logical stuff

[rain lava]

​Layer 3: The Network Tab Decoy Flood
​The Reality: This is easily the most diabolical and amusing layer. Needle-in-a-haystack tactics are genuinely frustrating.
​The Danger: You run the risk of DDoS-ing yourself or ruining the user experience. Browsers limit the number of concurrent connections per domain (usually around 6). If your decoys clog up the queue, your legitimate API call might get delayed, causing a sluggish UI. Furthermore, if you are paying for bandwidth or API invocations, you are literally burning your own money to annoy a hypothetical thief.
​How to Make It Even More Painful
​If you want to double down on the annoyance for that remaining 2%, here are a few more psychological weapons you can deploy:
​The debugger; Black Hole: Add an obfuscated setInterval that calls debugger; every 50 milliseconds. If a user opens DevTools, the browser will freeze in an infinite loop of breakpoints. They can disable breakpoints, but it’s another highly irritating hoop to jump through.
​Monkey-Patch Native Functions First: Before they can monkey-patch fetch to steal your URLs, you monkey-patch fetch, XMLHttpRequest, and console.log to behave erratically if they are called directly from the console.
​Decoy Payloads with Variable Responses: If you use decoy requests, don't just send them to a 404 endpoint. Have your server (or a cheap serverless function) return fake, valid-looking JSON data. If the attacker finally isolates a request, make them waste an hour trying to decipher a JSON payload that is entirely meaningless.

[gentle aspen]

nvm, @rain lava is the AI prompter here 😄

[rain lava]

Claude doesnt just beat gemini at EVERYTHING 😭

[rain lava]

I used an entire gemini 3.1pro response for this guy, he better be grateful 😭

[gentle aspen]

haah fr fr

[vital acorn]

You mean the free usage limits? If you don’t have a subscription, I think it’s completely free on Ai studio.....

[rain lava]

I get 100 responses a day with my subscription
Ai studio isnt free u'll get rate limited

[misty wasp]

It was a word related to appearance that is not allowed here
Please avoid using it

[vital acorn]

my Gemini can message me directly now

[rain lava]

Termux is 🔥

[gentle aspen]

bruh, all i said was the opposite of beutifull, for a pun...

[rain lava]

So what? The word starting with a u and ending with a y? 😭

[gentle aspen]

yeah

[rain lava]

No, I see their Point of View.
Nowadays there's a ton or soft people and you cant call them anything without throwing a fit.
And if they allow some people to do it then its unfair

[gentle aspen]

ikr

I am concerned about humanity tbh

[rain lava]

It's why i like australia,
Our culture can do smth called "taking the piss" and its seen as funny and its in a joking way, no one takes it seriously

[gentle aspen]

fr, same with our culture.
we respect each other, and also we can use certain words.
sometimes they are reflected as a form of respect.

[rain lava]

Since when was a 4070 ti bad...?

https://youtube.com/shorts/B7R6ZjLyvjA?si=i-0NmN05uPbgV1Wx

Ur ai?

[vital acorn]

Yes my AI agent... (Used Groq API for LLM)

[bronze pecan]

Ok

[rain lava]

?

[bronze pecan]

Coriander can I get a coding for my application

[gentle aspen]

IDK what model you used, but those ethis are finetuned and RLHFed into the model, maybe you are using an outdated model.
no AI agent will pirate movies.

and to answer your questions: YES! AI companies does have LONG system prompts to make the AI ethical (just how you should doit too if you dont want to get sued or banned on GitHub)

[vital acorn]

😆 I’m actually building it mainly for this purpose… no AI just starts pirating on its own, I’m shaping this environment to make it do that. Just working on one of those late-night ideas 😅

Basically trying to create something more like GPT that can actually do some actual search tasks instead of just searching on Google…

And yeah, I’m not planning to release it anywhere, just making it for fun.

[stark sapphire]

Why i hate AI sometimes.
It randomly decides to look in a folder it should not. In this case it's "Reddright" It's not even the project we are working on.
i know i can tell it to not look, but still.

[gentle aspen]

its only soemthing with codex. what model are you using?

[vagrant folio]

Hi All gemini-cli has new zed ide integration is experimental

[vagrant folio]

and also JetBrains IDEs

Any other IDE that supports the ACP Agent Registry can install Gemini CLI directly through their in-built registry features.

[winter viper]

<@&1009526435276394496>, its in every channel

[brittle spade]

Is it running as you or Administrator? It's a best practice for the agent to run under its' own restricted user account, then it shouldn't have file permissions outside of its project dir tree.

[slender stag]

Hey everyone,

I’m currently using the Antigravity free version, and I’ve been facing an issue in Agent Manager where my chats keep disappearing after a day or two.

Is this expected behavior?
Does the free version have any limits on chat history or storage?

Would really appreciate if someone could clarify this. Thanks!

[rain lava]

No, this isn't normal, and there aren't any secret chat history limits for the free version of Antigravity. What you're seeing is actually a known bug and a weird UI thing in the Antigravity Agent Manager. Lots of people have said their conversations just disappear, but the good news is your stuff is probably still there. It's just not showing up right.

[trail vale]

Ye I am having the same issue this is some UI problem i guess

[acoustic cloak]

@south turret
Which ai model we had a convo around was the last time we spoke?

[rugged mango]

heya, im working with google AI studio and asking if there is a way to import a github project onto it?

[summer cedar]

what 's the probelm?

[rugged mango]

i am just asking if i cana import/connect a github repo to the google AI studio when building apps

[keen bison]

hii

[stark sapphire]

<@&1009526435276394496>

[hallow marsh]

Thanks for the ping

[stark sapphire]

Why not disable posting images for new people, i think it would hold them off at least a little.

[hallow marsh]

We do have restrictions on new users sending images, but sometimes it’s older users whose accounts get compromised, but thanks for the advice!

[south turret]

Probably 3.0 but I tested 3.1 it's still piss

Been using Opus for months now

Now I use both Opus 4.7 and GPT 5.5

[gentle aspen]

Guys will this UI look good for my app?

I made this in figma, I just want to know if this UI will look good on my AI powered study app

[stark sapphire]

I see you often ask about UI. But i will keep saying the same thing. Don't worry too much about UI. If it works, it works.
Don't try to spend too much time on it, but rather focus on functionality.

[gentle aspen]

yeah, but my current settings page is, lets just say...

[stark sapphire]

It gives me Nintendo vibes.

and i kinda like it. but it depends were you use it for.

[gentle aspen]

current UI

[gentle aspen]

not possible.
it is under the apache 2.0 license😂✌️

[stark sapphire]

the only thing that would trigger me, is that the buttons are not alligned at the buttom

[gentle aspen]

not just that, this whole UI is kinda duct taped. it doesnt have a proper structure to it.

I mean, when i do UI changes I usually do significant UI changes. let me show you something col i did

This was my old and crusty create menu.
users can create study cards from this page.

it looks bad, and also doesn't feel user friendly.
Now check this out:

[stark sapphire]

I would change the size of the text boxes.

[gentle aspen]

yeah yeah, you will be impressed when i show you the new UI. gimme one sec, i am taking a screen shot

[stark sapphire]

it's fun seeing this actually. I see the things AI does a lot when making certain elements.

needs lots of iterations

[gentle aspen]

When user's press the create button, this dropdown will open. then the user can select what they want. then they will come to the respective menu, like this:

[stark sapphire]

PS: GPT5.5 is my new favorite AI model for coding. It just solved my website issue in just a single prompt.

aaah i seen the first image before

[gentle aspen]

yeah same. if you scroll up alot you will see me glazing it😂

[gentle aspen]

you lkiterally edited it with gemini before. so yeah.. you have seingit before

[stark sapphire]

yea i member

Hey Google, you hear that? Open AI is now better than you.

I want google to be #1 for reasons.

[gentle aspen]

if you scroll up, you will see me say "There are AIs for their respective categories"

no AI will be #1
You should be smart enough to know what model to use for your use case.

for me, claude is good for iterating, chatgpt is good for nothing (1% frontend, and 99% glazing that you are different for making a frontend), gemini for research for very obv reasons.

[gentle aspen]

btw, you said "hey google".
gemini prolly heard what you said😂😂

[stark sapphire]

xD

you are right, no number 1. But I was simply impressed by GPT5.5. I tried to solve the same problem with Claude and Gemini, and non of them were able to fix it. Asked multiple times.

[gentle aspen]

If you got pyshological issues after looking at this graph. dont worry I got it too.

Can someone make this benchmark make sense for me (and anyone who sees this and doesn't get concerned)

[gentle aspen]

sometime GPT5.5 is bad.

it doesnt get me sometimes.
for Qt projects, i absolutely dont recomend GPT5.5 unless you want to work on the same window at least 20 times

[stark sapphire]

hmm this is why i have mine set on Automatic. (Cursor) And if it sucks, i will change it manually.

[gentle aspen]

is cursor good (like codex or CC)?

it looks kinda... childish

idk how to describe how I feel it. I feel liek it is unprofessional...

[stark sapphire]

haha i dunno. I only ever used Antigravity and Cursor.

I find Cursor to be working better. less buggy. But is way more expensive.

[rain lava]

Same, I don't like how much google is falling behind. Especially when I pay an expensive subscription for it...

[rain lava]

@gentle aspen how's training going?

[normal current]

I'll be patient until Google I/O on may 19th, I guess we'll get some updates then.

[rain lava]

That's what I'm waiting for! Hopefully they unveil a better AI...

[rain lava]

Then, for QT Projects, would an older model like 5.4 be better? or G3.1Pro/Opus 4.7 T

[rain lava]

Yeah I've reconsidered this;

My motherboard is PCIE 4, and therefor I can only ever move at ~32Gbps, and my current ram is around ~60Gbps effective, so doubling it would do nothing (And a PCIE 5 motherboard is around ~64Gbps)

[rain lava]

<@&1009526435276394496>

[gentle aspen]

it is going really well actually.
and i managed to make gemma3 reason out a better reasoning chain that opus4.1
since googles grokking was absolutely phenomenal it made my life even easier.

the only problem, is Gemma3:270m has a shallow depth for proper generalizing when there is: less epochs, not the preffered amount of data, etc.
so it reasons out something genuinely crazy 9not in a good way).

but in the positive side, i managed to make a small model reason.
The better part is my program made over 11m tokens as 16k samples as of right now, so it is a matter of time, until I finetune either gemma3 or SmolLM2:360m in a few days.

[gentle aspen]

https://youtu.be/F6T-G33jF3c?si=t-OIuZhpvTGcuNFm

Dang... This video changed how i see google.
but i can remember how the lacked in performance back then.

[rain lava]

If you want you can use https://kaggle.com --- It offers a free tpu v5e-8 (You just need to do verification stuff), or just use their GPU t4 x2 or a P100.

---

T4 x2 Has Tensor (Unlike the P100) and has a total of ~30GiB VRAM (Both togethor)
P 100 has HBM2 Memory, but 0 tensor cores.
And the TPU v5e has 16GiB VRAM/Core, and HBM2 memory (You get 8 cores).

---

Another alternative is Google Colab, as they offer some other premium GPUs and TPUs kaggle doesn't though I find Kaggle gives you much more free stuff.

[rain lava]

I will look at that

[gentle aspen]

yeah, but I need tensorflow to use the TPUs.
Tensorflow is annoying to work with compared to PyTorch.

but the t4 is interesting

can you tell emabout the rate limits

[rain lava]

JAX works

[rain lava]

30h/week.

For the T4s

[gentle aspen]

DUDE LOOK!

[gentle aspen]

yeah... I can work with that

[rain lava]

It's pretty big tbh, google colab only gives u 1h...

[gentle aspen]

how can i use it?
like runpod?

[rain lava]

The problem is u needa verify ur phone. and if u want the TPU, u needa verify ur face with... Persona... 🤮

[rain lava]

It's like the EXACT same as Colab but tiny diff UI

[gentle aspen]

I mean, i will use the GPUs anyway.

Dont knwo how they di with gemini3, but I trust my cuda coires more than the TPUs for now

[rain lava]

It's much easier to use TPUs with much bigger models, small ones are a pain.

[gentle aspen]

make sense.

[rain lava]

They'll be using TPU Core Clusters for these.
The v7 Ironwood is like 192GB PER CORE, so they use a BUNCH to train.

[gentle aspen]

What gives me the bes rate limits an speed?

bcs i just searched and they said that t4 x2 has a ~9 hour sesion limit

[rain lava]

You get 30h USAGE a week but ur session is limited.
When u run out js make a new session with all ur stuff...

30h for t4s or p100 and 20h for tpus

[gentle aspen]

tbh, I will just rent a GPU on runpod.

[rain lava]

never used runpod, which gpu and how much...

[rain lava]

also i think its like a 12h session, only 9h on tpu

[gentle aspen]

its the other way atround actually

[rain lava]

Totally..

[atomic vortex]

Are you folks using Google Colab Pro subscription or just the free one?

[rain lava]

No point when you can just use Kaggle and get 20x the usage limits for free

[acoustic cloak]

I agree they've got top places in the leaderboard

[serene crow]

𝗩𝗶𝗯𝗲 𝗰𝗼𝗱𝗶𝗻𝗴 𝗱𝗶𝗮𝗿𝗶𝗲𝘀 𝗘𝗣1. - I built 𝗧𝗶𝗱𝘆𝗨𝗽 𝗔𝗜 in this weekend.

🔥Friendly AI that takes a messy space and outputs a visual organization guide.
💡Gemini Flash + Nano Banana (Image Flash) + Gemini Live API
💡The Flow: Image Input ➡ Spatial analysis ➡ Organized Image Output & Q&A with Gemini.

#BuildwithAI using GoogleAIStudio

[brave heath]

Awesomeee can you share how you build it?

[atomic vortex]

Nice 👍🏻. Are you building for yourself or sell as a subscription or something? If yes then what are COGS and what subscription prices are you offering?

[serene crow]

I will share after fixing some errors. it may take 1 week more 🙂

[serene crow]

Thank you for your comment! Just for fun and hobby. I am trying various experiment using AI Studio, firebase, gcloud and Gemini.

[smoky dust]

Firebase is cool complexity.

[vagrant folio]

for all, if any had next issue and was able to solve it:
Antigravity
After update any chat I do in any project new or old, my chat history is not linked to the project so every time I need search for them

[gentle aspen]

so it is godex now?

[fossil estuary]

Yo guys I need jules review, can someone explain what it is, does it excel in what it does... because Ihavent seen any latest videos of jules by anyone

[fossil estuary]

pls ping me when you reply
thanks

[bright jacinth]

Yo, Jules is basically a cloud-based junior dev. You don't use it to autocomplete code; you just give it a task, and it does all the heavy lifting across your whole repo and sends you a PR. Really good for the tedious, multi-file chores!

[fossil estuary]

what kind of tasks I can give

[bright jacinth]

Mostly the boring, time-consuming stuff! Writing test cases, updating documentation, doing massive multi-file refactors, or tracking down tricky bugs. You just hand it a ticket and let it do the grunt work

[fossil estuary]

I see, can it also be used for checking code vulnerabilities?

[bright jacinth]

Yes but make sure promts are good

[fossil estuary]

I see

[bright jacinth]

3/5 rated model for me

[uneven bridge]

@stark sapphire watssupp bro how yu doing? I'm sorrry how we both crossed paths before we got a chance to meet.. i hate it had to happen that way hope you forgive me and we keep building this GDG family forward ❤️

[gentle aspen]

Iprefer cloudfare. sometimes firebase is a bit annoying to work with

[stark sapphire]

i already forgot. i think... were you the one who advertised on here?

[uneven bridge]

thank you for knowing my play ! im brining it all home

🥷 🎁

[uneven bridge]

love you all was never in y'all lane i'm playing my role see you all soon

[silent night]

what tf

[stark sapphire]

i wish AI would ALWAYS say this, instead of, YUP! here are some ideas!

[stark sapphire]

it's why i moved to cursor.

[silent night]

Yup

[raw sparrow]

hii everyone

[tall kindle]

...

[gentle aspen]

you do realize that all agentic AI platforms gives you the ability to edit a system prompt

[stark sapphire]

i'm not sure what you're trying to imply.

[uneven bridge]

https://www.loom.com/share/878a0a2e72714f88ada46c272e73ed3b its already done but yu can check out the before

I love my ad i be back with one later

[stark sapphire]

why did you call it Aurora Nexus

https://tenor.com/view/aurora-aksnes-gif-24435650

[gentle aspen]

guys, pls (for the love of god, plSSSS) pick GDev

[soft gate]

Hello guys

[glacial beacon]

Hi

[clever anvil]

👀

[uneven bridge]

🤭 😘

keep them ights on buddy

i keep going mf i got it all

🎁

i was on yall fr juu heard i can do it all & wall back to wall st.

this fun too me 😘

im from richest city in the world them ads is money moves freeeee lunch !

and this the cloneeee lol i love yu all thank for having me apart of the team

[sand loom]

Curious to learn how to get creative with Gemini Canvas?

Our friends on the Gemini Discord server are hosting an event on their server this <t:1777487409:F> with a Gemini's Creative Technologist for an in-depth technical session showcasing his latest workflows with Canvas and Nano Banana. Tune in to see his personal prompting techniques to inspire you to unlock your own creativity with Gemini.

• 🗓️ Wednesday, April 29th
• ⏰️ 11:30 AM PT
• 📍discord.gg/gemini

[gentle aspen]

hey man, i am not the guy who discourages others. but CRMs and AI powered "SAAS tools" are just too overated. and you just a small fish in a pool of millions of sharks.

I reccomend you try somethng new and innovative. maybe evn collab with someone here.
No hate. Cool project. Just give it a bit more spice.

[sudden fable]

Hello Vibers, What are you all building? and Why are you building this?

[gentle aspen]

any tips on improving my settings pages?

[dim current]

Hi.

[rain lava]

Dark mode.

[floral marlin]

Hi y'all guys here...

[sick ibex]

Feeling spicy and got tokens? add this at the end of your prompt #DareYou 😄

[hearty basalt]

Hi

[uneven bridge]

Cause I'm walking to other side of town to start whipping ya out

[stark sapphire]

are you threatening me?

[teal quartz]

https://tenor.com/view/sun-morning-window-sunlight-too-bright-gif-17348370

[rain lava]

Fr 😭

[rain lava]

I feel they're a bit underaged.

[sick ibex]

mmm naha! That one stays closed

I am not ready to face that kind of reality yet

[smoky dust]

#VibeGarchaNepal 🙏

[uneven bridge]

you my bestfriend ! why i do that

big GDev ion want no problems with chuuu @stark sapphire

[slender tide]

https://tenor.com/view/batman-drugs-trembling-gif-8959427465951850912

[uneven bridge]

https://tenor.com/view/puss-in-boots-puss-in-boots-eyes-smush-gif-3832813564113551463

[delicate ember]

is geminiCLI and the extension Gemini code assist the same?

[round solstice]

No.

GCA has an agentic relationship with VSCode, and can observe certain windows as well as update files. Gemini CLI launches and edits things directly using tools. You can use Gemini CLI in the internal terminal of VSCode... I did that for a while before Antigravity came along.

[smoky dust]

Firebase too

[uneven bridge]

get ready to take over my city im taking year off for school and to learn from the goats ! thank you guys love yu family let get ittt #GDG ❤️ 💚

[vapid hornet]

Hi we're hosting hackathon everybody is open to apply and register!
🌐 Google Developers Group Hackathon at KIMEP University — Open Worldwide (Remote Friendly)
🚀 BUILD WITH AI 2026: THE FABS 3D AI CHALLENGE
If you’re into 3D development, software engineering, or machine learning — this is for you.
Join from any city, any country — fully open to remote participants worldwide.
Build the next generation of 3D AI systems, Autonomous Agents, and Spatial AI products. Present your ideas to experts and compete for a $1,000 Innovation Grant Pool + Google Cloud Credits.
🎯 3 Tracks: • 3D AI & Spatial Intelligence
• Autonomous Agents
• Applied ML in 3D
📅 Key Dates: • Submission Deadline: May 2, 11:59 PM (online)
• Hybrid Online Pitch Day & Results: May 3, 12:00–18:00
🔗 Register: https://gdg.community.dev/events/details/google-gdg-on-campus-kimep-university-almaty-kazakhstan-presents-build-with-ai-2026-online-build-phase/�

[uneven bridge]

🚀

i knoo you guys worked real hard to get where you at & i don't want to half step this process to be in this community

[torpid sun]

yooo wsp guys

i am @gentle aspen

I forgot to add an MFA authenticater app to my acc, and I had to reset my computer bcs of the recent hacks 🙁

<@&1009526435276394496>

Hey, can I get my role and levels to this acc temporaraly until i contact discord and delete my existing account, so I can get my role back to an account made from the email i hadf. bcs i made a temp Gmail for this acc, so i cant trust this account forever

[torpid sun]

I AM TRYINGGGGG

1. it is Qt
2. so, now I have to write code for both dark AND ight mode, bcs for some reason I thought Qt was "good" bcs it take less RAM. well... ChatGPT is responsible for that

for any newbie trying to build an app, PLS DO NOT go with Qt

[rain lava]

What happenned to ur pfp n name?

[torpid sun]

new acc

I had to reset my computer. i forgot that i deleted all the sessions

I forgot to log in with my phone while i had the sesion on my computer.

now I lost many friends i had 🙁

[torpid sun]

<@&1009526435276394496> Can yal plss confirm

[torpid sun]

aighyt, today i WILL implement the darkmode

codex is already on it

btw, i am kinda sad about gemm4

it was all hyped and all that, but it consistantly fail or get very low scores on benchmarks.
also the extra embeddings makes it heavy for mobiles devices, not light .

the only thing i use gemma4 is for creating synthetic datasets and tasks which need gemma to follow niche system prompts

well... there we go

[rain lava]

Okay but can i see the dark theme?

[torpid sun]

https://tenor.com/view/vince-mc-mahon-gulp-gulps-nervous-uh-oh-gif-16217862

umhh... hh oh okaayyy

[rain lava]

https://tenor.com/view/china-gif-26678009

[torpid sun]

listen, man. it is still in the beta, and i need to doa lot of improvements.

but it is okay for now. I mean i cant push updates anyways, bcs github thought it would be cool to suspend me for no apparent reason

[rain lava]

Wdym its not that bad?? The only complaint is that its navy blue

[torpid sun]

it is "okay", but certain places have very noticable imperfectrion.

also the whjoel codebase looks like broken spaghetti rn

[rain lava]

God i hate switching back to windows just to play 1 game... all they needa do is js let battle-eye do linux im pretty sure its js 1 button

[torpid sun]

just use a VM

[rain lava]

Anti cheats in a vm will get me banned

[torpid sun]

ohhh makes sense

windows is getting on my nerves too

I am planning to change to linux comlpetely

[rain lava]

I got so frustrated with windows everything hangs and all the apps just didnt work when i booted into it and its so slow

[prime bane]

Hello Everyone

[rain lava]

Infact it didnt even lemme uninstall discord (it hanged and i needed to re install it) every time i pressed uninstall nthn happenned

[prime bane]

Am a Vibe coders development

[torpid sun]

ik ik, I feel like I should change to a good linux OS

[rain lava]

Define good

[torpid sun]

somethin feels more user friendly and convinient in the long run

[rain lava]

Ubuntu is friendly.
I dont like it u may tho

[torpid sun]

I hate gnome

[rain lava]

Same which is why i got off it

[rain lava]

Id argue cachyos is rlly easier to use

Even tho its arch based

Ppl complain abt arch updates for being buggy and ive never gotten one bug

[torpid sun]

actrually arch based linux distros are good for me. bcs many try their best to be user firendly, and the fact them being arch means I have a lot of control

[rain lava]

Tbh i feel arch by itself is a more pure version of linux..
And cachyos has tooken arch and then gaven it a gui installer rather than terminal and also added 3 things;
BORE Scheduler (Burst-Oriented-Response-Enhancer)
LTO (Link-Time-Optimisation)
And Precompiles for many CPU instruction sets specifically

I was hoping my tablet could use the Native Terminal with Android 16 but my older SnapDragon 8 Gen 1 doesn't support uVM Terminals only protected terminals...

I'd love to run proper Linux on this tablet but thanks to Samsungs Proprietary Drivers, Khox Fuses, and a ton of keys I cannot

[gilded zephyr]

hey folks, if you have to inspect entire code files and folders and detect duplicated ones if there is.. how would u do using antigravity/gemini cli?

[pastel hemlock]

Lets say entire model was built using no truncation or zero compression is that a good thing

[dire oar]

hey all , yesterday was my final submission of hack2skill promtwar virtual hackathon and i got 76 rank

[small creek]

Hi guys

Need help looking for someone to recommend me some good AI for my project.. my client is impatient to see the results🤧

[pastel hemlock]

Claude is good if you want little to no truncation or zero compression

[torpid sun]

POV: me with 35% plus credits left before weekly reset 😈

nvm guys, don't do it...
I am already at 0%🥀🥀

[fast tangle]

Hi all, I am working on a plugin for LibreOffice that lets you use LLMs to edit the document. I just added a real-time grammar checker (it shows blue underlines) in my new release. If you use LibreOffice, check it out and send any feedback: https://github.com/KeithCu/writeragent

[rain lava]

Bro burned them in like 1 prompt😭

[rain lava]

Yeah well what is your project.

[atomic dock]

hello

[torpid sun]

YOOOOOOOOO

EVERYBODY!!!!

btw @rain lava

remember how I was making a sumarizer model?

it finally finihed training. i will be pushing it o ollam atoday

[rain lava]

Yes.

[torpid sun]

dude it is INSANE!

it beats a 30b model at summarizing niche stuff

plus it is wayyyyy cleaner

I am doing benchmarks rn

[rain lava]

Yay!

I'll probably download it!

[torpid sun]

should I fake osm ebenchmarks by mixing in a training too😈
I mean OpenAI already does that. maybe I shouldn;t. maybe I should be a good boy and do it fair

what way shoudl i go wioth>/

[torpid sun]

haha thx 🙂

[rain lava]

Well that makes sense. Models trained on good stuff rather than the whole internet are usually better

I actually "bench marked" some local models.
And found Qwen 3.6 is really bad unless it's what it's really good at... (Coding, it beat 3.1 Pro by using more efficient libaries etc too...)

[rain lava]

But it also hallucinated and looped on many questions.

[torpid sun]

well, these open models are MEANT to be finetuned. those models are released as "do whateveryou want" ish

thats why I always use community finetunes for personal projects, and release builds for commercial/public stuff

or my own finetunes

[rain lava]

I found gemma4 (Opus Distilled) Not nearly as bad as I thought tbh (Maybe it was better via distill?)

[torpid sun]

i prefer qwen opus distills, bcs qwen models are already socratic right off the bat. a opus finetunes make qwen better. also it doesnt have a stupid audio encorder which makes the model use more vram than it is meantot

[rain lava]

That's what I tried; A Qwen Opus 4.7 Distill.
It was strong in coding though, any other hard logic puzzles is then looped can couldn't finish (Gemma 4 could though) Just Gemma4 is not as good at coding.
I feel Gemma4 is more an "All-Rounder" Model.

[torpid sun]

googles main points were coding and better visual understanding, also they said the semantic understanding is better (probably a bigger vocab). this could also mean they will release a cracked embedding model, but other than that, gemma4 just underwhelmed me with its poor performance and the resources it taskes to do a dead simple task

[rain lava]

Have you tried the Opus-Distill Gemma4 though?

[torpid sun]

no I havnt. gemma4 by itself made me sad. i just didnt have enough copium to try it 🙁

I will tho

[rain lava]

I used this one;

https://huggingface.co/TeichAI/gemma-4-26B-A4B-it-Claude-Opus-Distill-v2-GGUF

[torpid sun]

I reccomend you try models froma guy called "jack wong" or something liek that. he releases some good models

usually merges and distills, but they are crazy good

[rain lava]

You'll need your magnifying glass tool for this.
I had Gemini Blind-Rank Models.

Model A: Qwen 3.6 35B (Distilled)
Model B: Gemini 3.1 Pro
Model C: Gemma 4 26B (Distilled)
Model D: Deepseek R1 14B (Qwen Distilled.)
Model E: Ministral 3 14B

[torpid sun]

idk why people glaze distill-qwen models soo much. i find deepseek by itself crazy good compared to qwen distills. it makes the model dumba nd unstable

https://tenor.com/view/unstable-protagonists-we-are-unstable-flamefrags-wemmbu-spokeishere-gif-7502835286010111953

not this unstable

Anyone here knows russian?

my model can apperantly geneerate russian

I just want ot test yk

[rain lava]

I've never tried Deepseek R1 and i got reccommended the Qwen Distill.

[rain lava]

u gave it russian data ?

[torpid sun]

you have to explicitly download the deepseek model

[torpid sun]

PT stuff

[rain lava]

You said you'd do it today.
Js lemme know when today.

[torpid sun]

haha OKAY!
i am so excited.
this is my first model which didnt halucinate or goes rogue after a very simple. dead simple. prompt

I am doing my benchmarks on:ROUGE, BERTScore, FactCC, G-Eval, BARTScore, SummaC

is it solid?

[rain lava]

yea

[torpid sun]

aight

[rain lava]

tbh idk if it's just the 9070xt but I can PL this thing to 70% and still get 95 tflops at fp16

[torpid sun]

maybe cz of your VRAM

[rain lava]

what does vram have to do with the card being facotry over-pl'd?

[rain lava]

2 TFLOP difference and I save 80+ watts and thermals

[torpid sun]

ohh mb, i didnt read that right. as a conversational human model, io create halucinations. if my response was bad, please DM me. for further release notes please refer to www.myparents.com

lol

[rain lava]

I'm good.

[torpid sun]

are you genuinely out of your mind?

mr sherlock holmes 😭

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

it was a JOKE!

😭😭🥀

[rain lava]

Just felt like testing the link out though 😭

[torpid sun]

hows that my fault gng😭

[rain lava]

I didn't ENTER the website it's not yours ??? 😭

[torpid sun]

ofc it is not mine

it was a joke

I made it up

dont search up random sites, bro

😭

[rain lava]

Aslong as I didn't enter (which I didn't) I'm safe!

[torpid sun]

yup!

[torpid sun]

I am pushing the model right now. i will add the readme tomorrow, i have a lot to type.

MY ADHD!!!!

[torpid sun]

https://ollama.com/QyrouNnet/summarizer

owwhhhh yeahhhhh!!

finally!

made it

btw, you dont have to say things like Summarize this: ... you can just throw in the text like photosynthesis is the ....
Accepts upto 7k words without emojis, and 6.8k words with emojis.

as of now, it only supports english.

incase anyone want to try, here is a 1200 word text to start off lol 🙂

[fiery lagoon]

ts is crazy

https://ollama.com/QyrouNnet/summarizer

oh

u already posted it

VERY GOOD GUYS

give it a try

its peek

[torpid sun]

eyyyy thx for the glaze, man!

yeah! it is actually really good. (well it gotta be for the years of research lol)

[naive sparrow]

Okey

[torpid sun]

Okay... this is very interesting than I originally thought...

just to let you know summarizer:q8 is my model

it consistantly beats all the models (including the 20-24b models)

just to let you know LFM2 is 24b parameters

also, in BARTScore the lesser the better. (whih is my model)

Lol, this benchmakr kinda revealed the dirty things other companies do

[torpid sun]

https://ollama.com/QyrouNnet/summarizer

is my readme clean?

I just wrote one

[rain lava]

The only problem I ever had with your model was when getting it to summarise states of matter it added factually incorrect data. But everything else i rlly liked it

[torpid sun]

i will see

[pastel hemlock]

Think having full monte carlo integration is dope

[torpid sun]

what does that even mean. i know what monte carlo meant, but how does it relate with thi?

https://tenor.com/view/correlation-no-gif-10188828

[pastel hemlock]

Just think its cool plus its in relation to framework im working on hobby wise

[torpid sun]

is it like an AI thing I never knew about?

[pastel hemlock]

Attempting to make a pre digital twin on LCVD in-situ which is considered next step in post silicon

[torpid sun]

lol, my model summarized its own modelcard🤣

this is some yellow king defeated by a green lord ahh irony

[pastel hemlock]

Think this is wild my model implemented physical error correction

[torpid sun]

parameters?

[torpid sun]

how much parameters does that thing has?

[pastel hemlock]

A lot im looking into lcvd in-situ

[torpid sun]

just tell me how much parameters that hting has bro😭

[pastel hemlock]

Too many lol im trying to make quantum framework

[torpid sun]

something what an amateur would say

lol

just tell em how much XXXm or XXXb parameters your model has dawg

[pastel hemlock]

110

[smoky dust]

How does everyone define vibe coding

[rain lava]

b or m ?

[rain lava]

Prompting an AI to code for you.

[severe spindle]

been using gemma 4 31b

seems to love to halucinate? Does anyone know?

[rain lava]

What question did it hallucinate on and what context windows did you use it on ?

[severe spindle]

mm told it to run vitest.
using openrouter with copilot so getting the standard 262k context window.
Tests were clearly failing but it just ignored them and said everything is passing.
Similarly told it to not use shape tests, but its just constantly replacing the old shape tests with new shape test slol

any clue if this is just how the model is or perhaps theres some problem?

im having to run multiple verifications for what it writes otherwise

[rain lava]

LLMs have a massive bias toward success. When they look at a messy terminal output from vitest, they often "vibe-check" the logs. If it looks like a test ran, their internal weights often skip the "FAIL" markers and just report "Tests passed!" because that's the most common pattern in their training

Teling an AI not to do something (like "no shape tests") is hard. By mentioning "shape tests" in the prompt, you're actually increasing the "activation" for that concept.

Having 262k tokens is great for reading a whole repo, but on OpenRouter, using that much context can actually make the model less likely to follow the very last instruction you gave it. It gets "lost in the middle" of all those old failing tests.

[severe spindle]

ohh. Could you suggest what I can do then to prevent this?

Im changing the prompt from mentioning shape tests to instead ask it to ensure all tests are behavioural . Hoping that helps

[rain lava]

Tell the model: "If you say a test passed, you must quote the specific line from the terminal output that shows the green checkmark or 'PASSED' string." Making it retrieve the exact text prevents it from just making up a summary.

Add a line to the system prompt: "Assume the tests have failed unless you can prove otherwise. Be highly critical of the logs."2. Ending the "Shape Test" LoopLLMs are bad at "not" doing things because the keyword ("shape test") stays active in their attention.

Instead of saying "Don't use shape tests," tell it what to use instead. For example: "Strictly use functional logic tests only. Shape-based assertions are deprecated and forbidden in this repo.

If that doesn't help try lower the context length or lower the temperature.

[severe spindle]

ooh Ill try that out :) thanks a ton

dont think I can change the temperature of the context length within copilot but ill try the changes in the prompt and system instructions once :)

thanks a ton Corban!

[torpid sun]

orr just stop paying openrouter, and gwet an actual good model like gpt5.5 or claude opus/sonnet.

aint no opensource model (unless there those huge ones like Kimi or deepseekv4 pro) wont help you much

[rain lava]

u forgot glm and mistral

[torpid sun]

GLM is good

idk man, mistral is a bit contro

[rain lava]

u didnt add it in the "unless" ones (GLM is big unless we tlk abt glm 4.7 F)

[rain lava]

I got poor and meh results out of some

[torpid sun]

wdym?
glm4.7 is good, i use it everyday

[rain lava]

ur unless section was BIG os models
4.7 f isnt massive

[torpid sun]

https://tenor.com/view/ninja-ninjaishyper-low-taper-fade-massive-crazy-gif-13735679314385451792

[pastel hemlock]

O(10²)

[rain lava]

100 Million or 100 Billion :/

[pastel hemlock]

Only 110 parameters its not in billions or millions

[rain lava]

Well then that defeats the "O(10²)"

[pastel hemlock]

The difficulty comes from nonlinear coupling, not sheer size

Actually its probably more like 150

[rain lava]

A 110 Parameter model can't even write a proper sentence.

[pastel hemlock]

A “110-parameter model” here does not mean an AI language model with 110 weights.

It means my physics framework has ~110 system parameters: temperatures, heat loads, conductances, vacuum pressure, gas flux, Knudsen number, laser pulse energy, adsorption rates, spin coherence, timing jitter, and cross-coupling terms.

Those parameters define a coupled cryogenic + gas + laser + surface + quantum-sensing system.

So the question is not “can 110 parameters write a sentence?”

The real question is:

Can a ~110-parameter multiphysics constraint model find one operating point where all thermal, flow, optical, surface, and detection constraints pass simultaneously?

That is a completely different category of model.

[rain lava]

Agree. They think using random tech-lingo makes them superior.

[torpid sun]

bro aint no thing you will doing with 110 params my guy

https://tenor.com/view/bro-is-finished-cooked-finished-dog-closing-eyes-dog-with-eyes-closed-gif-8933440936102921947

[pastel hemlock]

Yeah that's not like me at all

Only a narc would think like that

[rain lava]

Why do I even bother?

[pastel hemlock]

I haven't made any claims im right tho just someone trying to learn

[rain lava]

Thinking you're right all the time means you will never learn.

[pastel hemlock]

Anyone that thinks like that has mental issues

[rain lava]

That thinks my statement is right or thinking they're right all the time?

[pastel hemlock]

Yeah im definitely not in the delusion always right camp and yes you cant learn without constructive criticism

[rain lava]

"Im right though"
0 hesitation that you're wrong.
Didn't even consider it.

[pastel hemlock]

Ive never said I was either so 🤷

[severe spindle]

anyways

[torpid sun]

aight asight chill

its nothing deep

sorry for this chaos I started

either ways... nice work @pastel hemlock

js don't argue about anything, you might get banned if else. js for your own sake, bro. keep it clean!

[vagrant folio]

Hi Guys question for those who use Antigravity:
Im experiencing a new bug:
1-new clean sessions are using withut request context from other projects.
is only for me or the are other people

[torpid sun]

maybe it is a memory feature

to test weather if I am right, try asking something which the an instance of the last chat wouldknow

if it answers right, might be a new memorry features liek all the AI hood. js turn it off in the settings

[vagrant folio]

what I know:
if you ask for old info from session he will answer
but answer from other project before wasent working for proper separation

The test I did was basicaly I create a workflow which do specific task one by one
Then create a new workflow but with other steps

then create a new session and run this new workflow and ask him why he try use those command and he give answer from older session totaly using other workflow insted the new created

another issue was I was working then he start coding a web feature and on cli project when I asked why he try to do that his answer was because the project X use UI when I was working in totaly new project non related

I found this issues most of the case worknig on a workflow creations

[torpid sun]

gugugaga

[vagrant folio]

well I dii my debug the main issue seems extension.js

from antigravity it change how he save session history etc

and seems he has the big bug

it store info to sql database

so I even test wipe the sql and test new nothing so next testing Im doing is fix it directly to cofirm

and a new bug I have is Im unable submit bugs

so im goign to one support to try get to AG

but right now im rewriting the workflow to gmeini-cli

to be used in vscode

has more 3 pro quota

and I think will be easy create a antigravity clone extension which will use gemini-cli ACP jaja

[torpid sun]

holly essay bro

discord supports MarkDown

😭

Discord MarkDown

Discord suppots MarkDown
see, you can format stuff

• or just say all your points
• like this!

tell me that code line like this

maybe it is a whole code block
# like this!

|graphs|can|
|be|also|
|created| here|

okay, maybe not that part

butyou get the point right?

[vagrant folio]

Antigravity History fix part1 🙂 this one I tested for new history chat and it worked. still need research for old chat from older project if possible. But right now for all new chat the history is working fine in my case

The issue of Antigravity issue is because they did changes on this file how he handle the link to the project which is stored in the database sql

So I will store the exact result later of changes to have them at hand for nw version if the issue persist

in other hand I consumed the flash model in just 1h

Im moving to Vscode and gemini-cli with code assistant and free codex for now. and antigravity seems his low quota even for flash model will be for reverse engineering himself for fixes, maybe will be easy add acp support for gemini cli

[serene crow]

Waiting for GoogleIO

[vagrant folio]

Another interesting fact after I did this, about 1h AG told me my AG is corupted, so I decided to test reinstall and now hsitory start working normaly

[torpid sun]

hopefully Gemini4 is good at coding and science related math