File: 78A7BF30-FCB7-446B-8840-5(...).jpg (1.29 MB, 3464x3464)

1.29 MB JPG

>>108682497
VibeBUMP

After resisting the urge to slop I finally installed Codex, pointed it at my local repo for a work project and told it what to do for a feature. It built it in minutes... VibeGODS, I kneel...

File: 1763663560222556.png (592 KB, 3840x2160)

592 KB PNG

5.5 is out

how do I set up a github
can I have it all private so people don't see my vibetism projects?
@4chanGPT

>>108682541
Create a GitHub account, verify your email, install Git or GitHub Desktop, then create a new repository and select Private before you create it. After that, clone it to your computer, put your project files in the folder, add a .gitignore for your engine/language so junk files do not get uploaded, commit your changes, and push them to GitHub. Yes, your projects can be private so random people cannot see your vibetism experiments; only you and collaborators you invite can access them. Your GitHub profile username may still be public, and public repos or public activity are visible, so keep repos private, avoid adding people you do not trust, do not upload passwords/API keys, and double-check the repository visibility before pushing anything embarrassing.

>>108682546
Just install GitHub cli and the clanker can manage all that incl creation , setting privacy etc

>>108682541
dont use github bro ask chatgpt how to self host forgejo

mythos saaaars... it's over. codex ate our lunch again

>>108682568
no no no not like this

>>108682568
sam shartman really pulled off the comeback with gpt-5.5
I did NOT see this one coming
the slimy yid probably used corporate API data for training despite claiming it's not used for training
guess that secures another half trillion USD funding round

>>108682659
Man what I wouldn't do to work for him, goddamit

>releases 5.5
>didn't reset usage
SAAAAM

5.5 is 2x the cost of 5.4 btw

>>108682497
I love inseixes inoived

>>108682704
idc

>>108682728
proof?

>>108682659
opencuck have been dabbing on brahminthropic since 5.4

>>108682733
i never hit limits

why have no one plugged llm into rinworld?

>>108682748
liar

>>108682704
It's 0.5x the cost of 5.4 because it oneshots everything while also somehow being a lot faster and efficent.

>>108682745
>brahminthropic
kek. yeah idk what they are doing, especially with claude code which seems to be just a buggy mess now.

>>108682774
Prove it, oneshot something right now

>>108682821
Way ahead of you, boy.
See >>108682052
why do you think I'm shilling this aggressively? I'm not getting paid for this, I'm just acoustic

>>108682858
oh no no no this is bad, claude bros our response?

>>108682866
https://m.youtube.com/watch?v=a0jOZ-IMDNU

Where my vibe coders using local models?? What models, hardware we using?

>>108682937
I'm trying to make a dataset to finetune qwen 3.6 35b
I was making it with Qwen itself but I got tired of having the GPU fan running 24/7 so now I'm using GPT for the data gen

>>108682937
qwen2.5-coder:7b
8gb vram, 16gb ram
It's shockingly fast and occasionally good when it's not vomiting random Chinese.

>>108682950
chinese is good tho

>>108682950
uh uh fuck i have this hardware, i will look into setting things up, are you running windows or linux?

File: screenshot1741.jpg (627 KB, 1920x1018)

627 KB JPG

Had claude vibecode the road network editor and racing line calculation. Idk I had to handle hold it a bit though, seems like I cant just say "make the racing line system", it doesnt get it right.

>>108682985
Windows. I'm a lazy fat bastard and with hardware this weak it's not worth getting crazy with it. Easy route is install ollama, open powershell and "ollama pull qwen2.5-coder:7b ", head into VSCode and open Chat, under Model Selection click Other Models -> Manage Models -> Add Models -> Ollama. Now when you open the model list in chat you'll see qwen2.5-coder:7b, you can run it in Ask but not Agent. This requires a (free) Github Copilot account.

File: 1762076029948860.png (365 KB, 500x500)

365 KB PNG

>no longer have to continue.png
>now i can tell the model to RIP AND TEAR OUT useless code and it goes on a purge spree
thank you samuel h. altman

>>108682992
>racism in my /vcg/
What am I looking at here?

>Gemini is so bad that even Google is choosing to invest in Anthropic instead
If claude could generate images or textures, it would be perfect.

>>108683010
qrd?

>>108683088
5.5 is better than 5.4

File: 1767727009969905.png (198 KB, 581x751)

198 KB PNG

>>108683088
5.4 like many models was or became lazy and stopped early on tasks so you'd have to keep telling it to "continue"
5.5 isn't as lazy/is more autonomous so you can just tell it to work on a "long horizon" task and it'll go to work
for instance "rip (and tear) out part of a codebase (which is useless or anti-helpful)"

>>108683100
So far I'm not blown away. But I'm not gonna write it off or be negative about it. I think some of you guys are exaggerating though.

>>108683112
I have experienced much the opposite, but I am not complaining. It's getting more done and with more consistency, I'm happy to feed it a continue in exchange for fixing what 5.4 couldn't.

While debating whether the chatgpt pro version was worth it to test 5.5, I got an offer for one month of 'plus' free, so I got that.

The pro version was interesting me because I had read that there was a substantial difference between 5.5 thinking and pro.

But does that make any difference in codex, since even on the plus tier codex let's you choose 5.5 xhigh? Does pro offer one level beyond that?

Or as far as codex go, the difference between plus and pro tiers are strictly usage limits?

First time using codex.

>>108683189
AFAIK you don't get pro in codex with the pro plan. If you want pro for agentic coding you have to pay api prices - at least that was how it was a while ago.

Updated plan
/plan to preview
API Error: 529 Overloaded. This is a server-side issue, usually temporary — try again in a moment. If it persists, check
status.claude.com.
my disappointment is immeasurable and my weekend is ruined

>>108683195
grandpa I

>>108683195
Okay, so it makes no sense to get pro for a better codex experience, glad to have learned it before paying for it then. Thanks!

>>108683213
claude gives you pro if you get max

>>108683213
Well, no, it makes sense if you want way more usage. And also with the $20 plan the client cripples you to 200k context but you still can use the real 1m model with opencode.
And >>108683202 made me doubt, now I think I might have seen the pro models show up on the list the few months I paid for pro, I'm not sure. But I doubt it makes much difference, since pro is more for research, and the codex variants are specifically made for agentic coding.

>>108683189
OpenAI says 5.5-Pro is meant for shit like research and that regular 5.5 is faster and generally more appropriate for coding, they don't actually recommend it for coding and didn''t even bother to benchmark its coding ability.

File: chinese vs chinese.jpg (70 KB, 1200x630)

70 KB JPG

Has anyone tested how well Kimi K2.6 does in coding tasks compared to GLM-5.1? Especially in Go, Elixir, and React

I have an annual sub to Z.Ai Coding Plan Pro and I recently tried out Kimi K2.6 via OpenCode Go

Had both do a detailed code base review of 2 existing projects and Kimi K2.6 reported a lot more findings than GLM-5.1. I don't know how much I can really trust it's findings though. But if it's as good or slightly better than GLM-5.1, I'll put it to use

>>108683257
Z-ai had for months a backend bug/quantization issue where inference suddenly became completely broken after 90k of context use. I don't know if they fixed it but that combined with the price hikes made me completely uninterested.
>inb4 retards insisting "every model has it" and it was just context degradation

>>108683257
K2.6 overthinks, it's both faster and slower than GLM because it wastes so much time fapping. Their overall performance is similar, but if I had to choose one it'd be GLM.

>>108683257
Tell me more about K2.6

>>108683271
It's fixed at least since GLM-5 or maybe 5.1. I remember on GLM-4.7 it would just schizo out

also I paid $160 for my annual Pro plan. They are forcing older sub users to their weekly limits but I doubt I'd hit it. But OpenCode Go + Zen is my backup

>>108683257
I forgot to ask about MiMo-V2.5-Pro
It also scored around the same as Kimi and it's cheaper than Kimi K2.6

>>108683257
For my code DeepSeek V4 reported findings GLM-5.1 wasn't able to find. V4 also understood one of my quickfixes that is seeming broken was a response to server side changes out of my control. Though it also hallucinated an error so...

>>108683364
All these value chinese open weight models all catching up to a similar level are going to cause decision paralysis at this point

might have to come up with my own prompt benchmark to see which one does a better job

>>108683397
isn't the whole point of the chinks that they provide 80% of the value at 5% of the cost? simply multiplex and check yourself which you find better

>>108683471
Still true of course, but DeepSeek's prices went up 8x with v4.

I keep hitting my limit with claude pro. Do I really have to pay €180 + 25% VAT per month to vibe code without hitting a limit when doing normal work? or are there better and cheaper options than claude.

>>108683800
what if we all made an llc and pretended to be a legitimate business for better pricing

>>108683800
the new Codex hits just as hard but no one has made a solid comparison of token usage across platforms so I’d say just stick with Claude.

>>108683808
I might have to do that. There is also food that I want to buy but only companies can buy it. They dont sell the same stuff at supermarkets. The closest thing at supermarkets is worse and also much more expensive.

Why the fuck do companies have these privileges REEEEE.

>>108683800
Yes there are
Claude is only for Max

>>108683835
such as?

>>108683811
You don't need a comparison if you've used both
Claude has always been more stingy

>>108683841
Codex, Kimi, Opencode Zen, Minimax

>>108683853
Are those as good as opus 4.7? im a vibe coding newbie

>>108683864
No

5.5 remind me of 4.1, they don't really think so they keep correcting themselves in output

I should probably switch to Codex for my exoplanet app, but I want to congratulate Claude at the end.

>>108684114
you can just gaslight it, simply append the claude catchprhase to all your codex commits

File: froggg.jpg (8 KB, 229x220)

8 KB JPG

>check local tech forums
>they are buying codex plus from 3rd party broker at 1$
goddamn it why can't thirdies keep things simple

>>108684283
Not having first-world levels of money is the mother of at least some invention

>>108684283
what?

i never tried it with 5.4 but 5.5 is suprisingly good at webshittery just by using python and playwright

whats a good beginner vibecode adventure?

>>108684574
- 90s-style website dedicated to your favorite 2000-or-later anime character
- CLI-based video game (not sure what’s fun, maybe a Wordle clone)
get used to using Git, too

File: Screenshot_20260425_020254.png (63 KB, 1702x247)

63 KB PNG

>>108683471
I just did actually

I'm still going to run more tests, but I have an identical prompt to all 4 of these models to make a Go + react media sharing app (Pomf clone with Accounts) and to have image/audio/video preview pages, custom player controls, user dashboard and a mock billing system

I gave each model the same lineup of subagents to work with like I normally do with GLM-5.1 in OpenCode

Kimi K2.6 - Best Frontend. Cheapest but probably cause of the OpenCode 3x usage limit may affect actual "cost" on their usage chart

MiMi V2.5 Pro - Best Backend, by far the fastest finishing in a third of the time. Technically the cheapest

GLM-5.1 - Buggy Backend and encoding. Frontend design was completely dogshit trash I don't know what happened. I've gotten good designs out of it most of the time I used it, but this was fucked. It didn't know what to do with the theme

Deepseek - API error towards the end and could not fully finish. The registration was broken so could not easily test what everything else looked like. Frontend was trash but not as fucked as GLM-5.1

>>108684644
>pomf clone
speedrunning getting raided by glowies?

>>108684649
I figured it would be a good test and easy to describe

I also used to run the biggest Pomf instance. I don't anymore cause it was getting on the verge of attracting glowies

Are we at the point where I can just vibe-create my game?
Including 3d models

File: 1763318651884503.jpg (194 KB, 1448x1086)

194 KB JPG

Accelerate.

>>108684702
I wish the y axis was some indication of ability rather than literally the release number

>>108684717
GPT-9.11 will beat GPT-9.8

>>108684574

Hmm
If i was you i would go for some flutter app. It's just damn easy to create a functional app with flutter ready to use and use it.
Maybe ask for a terminal with advanced features with gemini-cli.
When you'll start using your app on your computer the dopamine hit will makes you want to just add more and more and you'll notice it's limitations.

>>108684644
2nd test was a simple static site design for a dev platform

This time, GLM-5.1 was the best design with many intricate sections, MiMo V2.5 Pro Runner up but a bland hero but consistent style, then Deepseek V4 Pro, and Kimi K2.6 was last place

Time to complete:
GLM-5.1 - 17.5 mins
MiMo V2.5 Pro - 8 mins
Deepseek V4 Pro - 7.5 mins
Kimi K2.6 - 18 mins


A static HTML design is simplistic to do, but the difference in design here is opposite of the pomf clone design. However they didn't have a theme to adhere to

Holy moly guys, has my prompt game dropped so much or did claude fees goddamned skyrocket? Agents feel slower, too.

>>108684913
I’m not sure if I’m confusing it with the new ChatGPT model but I think the new Claude does take a bit longer, but it’s generally better
also the new default for Claude is xhigh and max takes like double the tokens for only marginal benefit
make sure you’re only doing xhigh for most things and only deliberately choosing max if you REALLY need it

File: 69e1088ee4f61fecf4983993_(...).png (78 KB, 1920x1080)

78 KB PNG

>>108684913
https://claude.com/blog/best-practices-for-using-claude-opus-4-7-with-claude-code
really study pic related

File: coopprinciples-e154023428(...).jpg (104 KB, 1024x527)

104 KB JPG

>>108683808
>>108683824
think correct term is co-op, i looked into this a bit awhile ago and it sounded like a great idea but i completely forgot about it. I'm down for food/ai/other perks marigoldwhale@proton.me

>>108684949
lol

MiMO is pretty good... you guys are sleeping on it

>>108683808
are the rates actually better for enterprise vs individual? as far as I can tell the subscription pricing has been the better deal whereas api get gouged

File: 1775759175818611.png (289 KB, 352x910)

289 KB PNG

Are there any models other than GLM 5.1 and MiMo V2.5 that do this

File: coom.gif (134 KB, 800x800)

134 KB GIF

Does anyone here use cursor or github copilot with local models? I know LM Studio can do API output so its technically possible, but just reading around on google details seem scarce and debated.

>>108685159
I've been trying qwen3.6 27b, struggles with using custom mcp servers though. had opus run setup.

File: big daz.jpg (31 KB, 500x500)

31 KB JPG

>>108685159
oi cunt what you got against cumbrians eh?

>>108685159
man this gif changed my life the first time I clicked it, legitimately made me actually think about things and get up and do shit for once and ended up with 4 years in prison

>>108685209
nigga wtf

>>108685159
>yogapants in thighboots
hnnnnng

any richfags using amp or droid
are they really the supercars of the tui world

>>108682541
kys

File: 1770756611090410.png (1.51 MB, 3625x4021)

1.51 MB PNG

>Trying to burn through my tokens as fast as I can with Codex 5.5 xHigh x1.5 since I'm going to be gone for a week
>Thinking of as many fixes/features as I can
>It's nailing them all
I'm gonna be honest OpenAI nailed this model. 5.5 is amazing and I wish I had more time to fuck around with it. Anyways, cya guys!

just realised both ant and openai have done like 6 releases since gemini 3 was release
wtf is google doing

>>108685543
Google is probably the only one near profitable at the cost of being terrible at doing anything. Still trying to understand what they want people to do with antigravity when the usage gets eaten up within 3 prompts.

File: file.png (16 KB, 525x263)

16 KB PNG

>>108685604
i just realised that they're still too afraid to take the preview label off their models
flash 3 has been out since mid december and has had zero updates
they're going to roll up to i/o shit out 3.5 and then do nothing for another 6 months aren't they

>>108682497
Vibe coding is so easy and fun, but im stuck using free 5.3 Codex in Cursor.

File: HGvB2GEbcAA66ip.jpg (157 KB, 795x1237)

157 KB JPG

I've been using a mix of Claude Code, Gemini CLI, and Codex to improve open source tools with which I plan to replace them, because I kept getting ratelimitcucked by them working on my actual job and hobby projects. Now I just need to assemble my Framework Desktop and see what I can run on 128GB with Ollama hooked in to my local open source stack.

>>108685154
Idk, but you can tell other models to do that.

why's opus the only model that doesn't shit itself working with frida

File: HGry4RhWgAA9YGI.jpg (61 KB, 1048x312)

61 KB JPG

kek opussy4.7

>>108686384
AGI in gaslighting

>>108685543
I hear Nano Banana is pretty good

>>108686480
It's shit compared to GPT-Image-2

>>108686484
The thing that OpenAI just released?

I find Gemini 3.1 to be pretty good for the price.

File: 1777120029040540.png (33 KB, 804x655)

33 KB PNG

I posted my website and people said it was slow so I've been optimizing it.
It's a nextjs project on turbopack and for some reason it was loading the entire lucide-icons library instead of only loading the icons that were actually being used. Also a lot of useless zod schemas were being used and I also optimized the thumbnails.
Now I'm changing the rendering strategy.
https://umigalaxy.com/explore/general/348-this-site-is-slow-as-hell

File: 83d014f19cf.png (173 KB, 512x386)

173 KB PNG

Github copilot's poor tier subscription does not include gpt 5.5, it's over for us.

anyone tried subbing to one of these things from like turkey or india or sum shit.
search results from around a year ago suggest it was working, but seems to have gone dead since.
i wouldn't mind gambling on a second plus account.

>>108686731
I just tried from an Indian server on my vpn and it gave me the pricing in Singapore dollars which is more expensive than the plans in the US dollar kek

>>108686787
lmao. i think the route people were using was the apple appstore.
set region to whatever > load giftcard > purchase sub via apple pay to obscure payment method
i think it's usually done for spotify and youtube subs, but they are starting to clamp down

does cursor just use API pricing?
for GPT 5.4 for example.
I'm wondering if I should go dual subscription with Codex + Cursor, or is Cursor actually completely worthless when you could simply pay for credits from OpenAI to get the same result?

qwen 3.6 plus fucking sucks ass. I'm never falling for the open source model meme again. And the api costs suck.

I want to build my own multi model orchestration plugin with subscription based agents, just by copy pasting messages between different terminals, and reading the agent logs. How retarded is this? Will I get banned for this? How do they usually detect this?

>>108686956
you can go much further than that tb h.
with the exception of google and anthropic, the others will usually just let you use their auth in a different harness (like pi) and you can do whatever you want.
google will force you to go through gemini cli via acp and that's explicitly allowed.
anthropic is where i'm not sure - there's an agents sdk in cc but you never know with them.
pi + nicobailon's subagents/intercom extensions might be a good place to start.

>>108687012
Ok, that's good to know, so Codex at least would be covered. I might also check out some models I haven't used yet to integrate with pi. Thanks.

>>108687157
pi and opencode are officially approved by openai.
you can just use use your sub in pi and spawn as many gpt subagents as your heart desires.

>>108687218
why use pi instead of codex?

File: HGtRmzAaEAEMEV7.jpg (583 KB, 2898x1708)

583 KB JPG

>>108687351
easier to extend so you can do dumb shit with it
or build the ultimate corpo experience

>>108685694
They still won't even give me access to 3.1 on the code assist standard plan.