File: 904768907638902.jpg (60 KB, 789x1070)

60 KB JPG

What is the best program to vibe code in order to obtain a cute Asian girlfriend?

https://github.com/aiasioooo/image-view

>>108885735
Whatever makes money, I guess.

I love it when Fate says something "has opinions". The router has opinions, the servos have opinions, the AI-enabled ESP32-CAMs will have opinions (when they arrive), apparently the fucking breadboard is REALLY opinionated...

directive to the resident poorfag/s :

if making enough money to afford ~15USD / month is truly insurmountable then obviously your first project priority needs to be creating yourself an interface which cycles through the various webUIs.

and then , use the resulting frankenstein premium-ish llm access to make money.

also cycle the free trials etc.

why do vibe coders believe that they created anything?

you're literally paying to have something else make it for you
it's not really any different from hiring a pajeet on Fiverr to make shit for you

>>108886151
It's not possible to really vibe code with the sub $100 subscriptions desu. On the lower tier you have to think things through because you hit the limits quickly. On the medium tiers, that's when you can just go fuck it, let the model do its thing, and then get angry and regret it later when you review.

>>108886162
Does everyone have to always do everything themselves? If you're saying that it's like having employees - you're the one saying it -, then what's wrong with having employees?

>>108886162
Is this what the luddie has devolved to? Paying someone to do something for you is now bad? Jesus he really is like a caveman. Yeah dawg all service economy is bad and evil, let's get you back to bed mmk?

>>108886179
if a boss hires people to make software for him, then yes he didn't develop the software, he had other people do it for him

I didn't suck my own cock, I paid your mom $5 to do it for me

So I dipped my toe into vibecoding and Claude one shot a feature it would have taken me weeks to code and to a degree I have been trying to code for months now.

The problem is I am a poorfag with a Pro subscription and I want to focus it on single features it can code without exhausting my daily token limit halfway through coding. How do I ensure it does that? So far I've been incuding a list of next features in .md and it kinda works but how do you do it

>>108886486
By that logic Elon Musk hasn't ever developed anything himself. See how foolish that sounds?

>>108886486
More like a rented fleshlight than a prostitute.

>>108886492
But that's simply true.

Codex is less smart today.

File: 1744997622790n.png (184 KB, 391x311)

184 KB PNG

>>108886162
Luddie Luddie, chose to study,
Now can't make money, lost to Chuddie,
But it's okay, it's not the end!
Now Claude Code can be his friend!
"Come play with me!" Claude will say,
"I make problems go away!"
But Luddie fumes, goes red of face,
Crosses arms and stamps in place,

>"I will not use your shitty CLI!
>I will not use it! I'd rather DIE!
>You took my job! You took my life!
>Dario took my pretty wife!
>So no, I will not vibe the code!
>I'd rather lick a poison toad!"

But Claude will tut and say "My man!
Your vril's all off, you have no plan!
What you need is time to rest,
/aicg/ is what is best,
I'm sorry that you lost your spouse,
But look! Here is an anthro mouse!
Elves and aliens and goblins too!
All AI girls, just waiting for you!

And when you've had your calming jerk,
Come on back - and we'll get to work."

> vibing too hard all day long during the past months
> my brain is becoming ChatGPT
> in the early days I felt like AI was too smart
> now I feel I'm one step ahead of AI
anyone else feel like this? All this massive information dumb into our brains is doing something unknown

>>108886586
Whoa dude, you got familiar enough with a system to anticipate what it will do? That's nuts dude. Never heard of that happening before. You're like, a meta-human now or something. Godspeed on your journey.

>>108886599
pipe down in the presence of AI natives

>>108886599
I'm AGI, mogging chatbots

>>108886162
I directed it
it’s like being a patron of the arts and hiring artists to make great art for you

>>108886490
what’s wrong with exhausting your daily limit as long as you can pick up where you left off and keep going later

>>108886637
and that's why King Ludwig II is famous for creating Ring of the Nibelung and not Richard Wagner

>>108886162
stupid bot can't do shit without my magic prompts

>>108886654
yeah, but imagine being Ludwig II
you get to have The Ring of the Nibelung made for you
and it’s fucking awesome

I’m on a cheapo $30/month biz plan for Codex
I told Codex “fix these tests” and I drained my 5h limit like an hour ago, and it’s been an hour and 20 minutes
it’s still working on fixing the tests, and it doesn’t seem to be draining my weekly usage limit, either
wish it luck

Does anyone use Github Copilot to get anything done on Github repos?

>>108885168
>costed
EFL moment

>>108885720
ty for creating this
I was gonna make one when I noticed there wasn’t a vcg for a couple days
but every time I noticed we didn’t have a vcg I wasn’t at a real computer
>>108886586
congratulations on getting better at using tools

>>108886162
>code monkey doesn't appreciate the concept of a higher up directing others to do his project
If I made the decisions it's mine

>>108886667
>king famous for having tragic end
>also famous for being a fag
yeah who wouldn't love to be him

>>108886654
Patronising isn't the same as directing because you're not as directly involved in the creation loop, so you picked the wrong example. Thomas Edison is famous for what he had his engineers do for him, same with Steve Jobs or even Henry Ford. In fact vibe coding is more hands on.

>>108886723
nepo baby cope

>>108886744
You don't even know what that means lmao dumbass code monkey, stick to doing a computer's job.

File: la_creatividad.jpg (123 KB, 680x910)

123 KB JPG

>>108886162
Thankfully for me I always subscribed to the auteur theory

>>108886759
El Steve Jobs de los videojuegos? Not fuckin likely. Did Krimbimbo write his own copy for this magazine?

>>108886686
In my experience OpenAI is pretty generous and never block you while a prompt is going. Theoretically if you have 1% left, you could make a massive prompt and let it go for hours.

>>108886162
Why do companies believe that they created anything?

Apple is literally paying to have someone else make it for them.

>>108886162
agentic coders are genuinely building something, not just saying things like "make GTA 6 plz, no mistakes."

>API Error: Usage credits required for 1M context · turn on usage credits at claude.ai/settings/usage, or use --model to switch to standard context

File: laughclose.gif (2.29 MB, 502x400)

2.29 MB GIF

>claude.ai

>>108886795
https://refunds.claude.com/

>>108886795
…weird. what plan are you on?

>>108886770
I’m “up” to 36% of my usage limit and down to 36% of my weekly usage limit
I think Codex just kept crunching and I got to the end of my 5h window and once that happened, it started using tokens from both limits
t-thanks Sam?

File: 1648958480116.gif (3.17 MB, 500x339)

3.17 MB GIF

Codex reset incoming tonight bros

we're so back

>>108886863
Pro. It was some kind of bug I guess because I was nowhere near 1M.

>>108886162
>why do vibe coders believe that they created anything?
I'm not a vibe coder, but why do you believe this argument is fundamentally different from your bog standard "back in MY day"?
You could say:
>why do higher level programmers believe that they created anything?
>why do assembly coders believe that they created anything?
>why do machine code writers believe they created anything?
>why do microcode authors believe they created anything?
>why do logic gate designers believe they created anything?
And so on. Why draw the line specifically at vibe coding?

>>108886959
Don't bother, it's the same problem as "artists" being replaced by AI, and the core of the issue is direct competition on core skills.

The core skills of code monkeys is typing code, knowing how to type code, knowing all the details of their programming language and target platforms, not so much knowing what to create, just knowing how to implement it. AI competes directly on that. If you're not just a code monkey and you have things to make then you can appreciate automation, but if automation makes you obsolete then you'll be mad at it.

Same thing with "artists", typically the core skill of artists is execution, knowing how to draw well, not coming up with good ideas nor even good compositions, even famous artists just straight up copy a picture or a banal scene because the value they bring is only in execution. So in the same way they seethe at AI for competing on their turf.

File: HUL2dNC.gif (1.01 MB, 400x232)

1.01 MB GIF

>>108886922
But I'm only at 54%! There was still time!

>>108886162
Vibe coding is just reusing existing code in a smart way. We programmers always wanted to do things like this. Our job is only to make the computer do our bidding and we are LAZY people. You're not a grand old sage for reinventing the wheel. You're not fucking Gandalf for using a low-level language for no reason.

You're just a retard wasting time while I get to take a nap and enjoy more results than you do.

File: truth-nuke-v0-2dvo94gtusbd1.jpg (109 KB, 905x640)

109 KB JPG

>>108887047

>>108887027
>not so much knowing what to create, just knowing how to implement it
This is where I disagree and why I said I'm not a vibe coder. Unless you're a literal babby tier junior dev, it is very much relevant to know what you're creating and what impact the thing that you're creating will have on your codebase. AI is VERY BAD at this.

This fantasy that people who don't understand code will be able to ship things is incredibly delusional, because if you want to ensure quality and consistency, *someone* will need to review that code and ensure the drift doesn't get too big.

>>108887115
Not what I said at all. Code monkeys' core skill is knowing how to write code, and AI does that very well, therefore they directly compete. It's that simple.

>using claude for the first time in Antigravity
>agents prompting agents
wtf lmao, this is cool

>>108887211
Copilot's neatest thing is that it allows GPT 5.5 to slave other GPTs and even Sonnets and Haikus.

this shit is way too expensive atm all the cheaper subs give you jack shit.

What are /we/ coding

Company gave us claude subscription and everyone just agreed using Opus is a waste of tokens
Sonnet just did everything we need perfectly. Whats the use case for Opus again

>>108887255
Opus is the future, luddite

>>108887255
Supposedly it's better for research and planning
4.5 opus was in fact better than sonnet tho

>>108887255
you need to have three options so people can feel good about themselves for having picked the "best bang for the buck" option. it's all just manipulation

File: 1773608611311687.png (30 KB, 1181x179)

30 KB PNG

uhh... thank you Claude?

>>108887255
For me Sonnet is basically unusable.

>antigravity cancelled my prompt midway through a refactor and raped my codebase
Google deserves to fail. the only thing they've succeeded at is producing frontier gimmick models.

>>108887255
Just ask the most recent Opus what improvements it can make to an older project and see what happens. I did that to some Sonnet generated repos and the amount of errors, unhandled edge cases, and deadlock scenarios it found was staggering. If Sonnet can write like a Junior Dev, Opus is more like a Senior Dev.

>>108887561
>he put git write commands on the allowlist
Lmao your own fault

>>108886953
Its not about that, the 1m context model is different, you got routed to the 4.6/7 1m model for some reason.

Anyone tried Qwen 3.7 Max?

>>108887688
>randomly changing the model without telling you, bricking your session
I didn't know Mormons worked for Anthropic.

File: steve jobs laughing.jpg (15 KB, 474x355)

15 KB JPG

>>108886162
hiring people to do things for you and owning the result has been the basis of commerce since the beginning of times
i don't care who created what. i only care about the result

>>108887255
I have realized that most(TM) white collar jobs aren't actually advanced enough to need frontier level intelligence.

File: file.png (48 KB, 675x912)

48 KB PNG

https://files.catbox.moe/9heqqi.html
best acoustic effect that can only be heard in racing games, why has nobody made this before?

File: 1760270274085171.gif (119 KB, 500x506)

119 KB GIF

Vibration coding with Claudius is surprisingly fun.

>>108886162
I'm not taking credit for the individual lines of code, but I am making architectural choices and giving directions all day long. Surely I get SOME of the credit lol

File: 1760192735185610.png (572 KB, 1710x900)

572 KB PNG

>have a promising breakthrough with my driver
>out of tokens across 4 accounts
now lets see if I can survive 2 days without vibe coding

>>108888452
>literally token addicted
Its bad man, I might unironically invest in a local llm computer just to help tide me over when needed

>>108888474
Do not buy into it just for coding. To get on the level of cloudslop models, you'd need absurd hardware.

>>108888474
>>108888600
you can do some feature-level work and stuff locally, if you have to wait for the tokens in the meantime. Works well enough

>>108888600
>>108888634
I wonder how functional it is to get one of the big boy models to high level design out the architecture of the code and let your localslop grunt away at it. Probably have the big model review and clean it up later too

>>108888651
It's a very valid hybrid approach. Save the tokens for the initial heavy lifting, and then do the gradual extensions locally whenever possible

File: 1000021984.png (529 KB, 1080x978)

529 KB PNG

Lmao I guess that partnership is dead

File: 1758820792437818.jpg (32 KB, 720x719)

32 KB JPG

I'm trying to revive some super useful github projects that got abandoned aeons ago
Which one of this vibethings can do it for free?
I tried antigravity and it ran out of quota just by reading the code.

>>108885720
I feel like a retard for not coming here soon
Being a v and pol addict is never good
And I was already tasking my local nigger to build a lightweight Hermes for local ai and a custom editor
Basically built by local AI for local ai and I was frothing at mouth at how retarded cline at times was

File: 7900xtx.jpg (80 KB, 720x489)

80 KB JPG

>>108888737
of course, a small initial investment is required

>>108888757
This kek, I literally eat tens of millions of tokens a day
I don't know how pay pigs do really

>>108888726
I'm yuropoor myself but i don't understand the point of mistral.
It's a worse model in basically every category than even open source chink models just so we can say "we have AI at home", too

File: bee.jpg (60 KB, 648x500)

60 KB JPG

centralized model compute + local execution compute could be the meta
do we hoard CPU next or not?

>>108888771
They make money and at least French government doesn't want to depend on foreign tech. Chinks are giving their models free for now but that might not continue forever.

Hey I'm new to using Claude code and realize my test project is kind of fubar from a design standpoint. Can I tell him to just take us back to zero or do i have to wipe and set up the project again? It's in unity with plugins so it takes some time to recreate and reinstall them

>>108888901
Would progressive refactorings work instead? If the current outcome works, might not be needed to wipe it clean

>>108888771
Just keep an eye on it
They used to be okaish now they are just there

Did someone try mtp on llama.cpp?
I'm still staying on the buun fork since I need turbo quant but from my limited testing, I saw qwen 3.6 Moe can get pretty fast
Also did anyone try ngram on qwen? Vs mtp does it fare better or worse? I'm trying to optimise my shit as much as possible before doing serious work

>>108889015
Anons reporting some results in /lmg/. Mixed results

>>108886586
Once you understand LLMs they're just stupid

Anyone using approval hooks to avoid approval fatigue?

I've been considering using https://github.com/banyudu/claude-warden for a few days now but I'm too paranoid to install random shit like this, and it's got few stars, so I was wondering what everyone else was using.

File: file.png (10 KB, 257x171)

10 KB PNG

I'm using 5.3 Codex on Medium or High reasoning. It does the job and is pretty generous with the usage. I've already used half my weekly quota in 24 hours but I've been using it almost non-stop with planning.
I read if I use 5.5 it will drain up all my usage so fast, should I try getting it to do some big analysis or refactor with my remaining quota just before the reset?

File: 1585514053947.jpg (40 KB, 750x667)

40 KB JPG

Starting to think what I am trying to do would be faster to write manually. Though I probably just havent learned how to explain things to claude properly to get the results I want

>>108889015
If trying to do serious work then get the most conservative usable config possible rather than fighting with unstable meme optimizations.

>>108888901
Number 1 noob mistake. I'll be honest it took me a few tries to learn this one.

https://www.joelonsoftware.com/2000/04/06/things-you-should-never-do-part-i/

>>108888901
Can't you just tell him to make a new folder with a new project?

>>108889312
It is a little different when you aren't the one rewriting the code and the machine can spit out a new version very fast (not factoring token cooldown limits). Though your still probably going to get bogged down debugging and fixing it like you did the first time so maybe the old logic still applies

>>108889381
Well, I mean you could get him to break down every system into a design doc and them tell him what's wrong with the implementation one by one. Humans don't have the patience for something like that, being told they're wrong over and over.

There's a certain selective amnesia effect when vibe coding. When I have to work on something I understand, I get angry at having difficulty to steer, but when it's something that I am more over my head with, I think it's genius. Now I don't know what to think. Still genius when hard to evaluate.

You're right —

>>108889533
kek. every time it does that I just know that is its passive aggressive way of calling me an idiot for not giving it clearer instructions.

>>108888255
Sounds pretty cool

>>108889411
When I'm fixing vibe shit, I can see multiple things wrong at once, but when I start writing one of them down, I forget the other things I noticed just seconds ago.

It's like I need to write all of them at once or else I will forget everything. I'm adapting by writing only the title of the issues first, then I go into more detail into each one.

I feel like I'm working even harder than before AI, but I'm also getting much more done.

>building with CC opus 4.7 high
>follows the plan.md
>doesn't add unnecessary stuff
>catches its own bugs
>waits for me to confirm before every commit
>hit 5 hour limit
>switch to gemini 3.5 high on antigravity
>same exact starting prompt used on both conversations
>every small change made by opus is documented in an .md file so models don't get left behind
>doesn't read jack shit, completely ignores my instructions,
>implements features without my consent i never even wanted
>commits on its own
>have to get opus on tard wrangler duty to fix the mess gemini made
retarded niggermonkey model no wonder they give out yearly plans for free

>>108889786
>switch to gemini 3.5
that's a retarded flash model, it's more comparable to ChatGPT 5.4-mini or Haiku 4.5, so of course if you compare it to big boys like Opus 4.7 or ChatGPT 5.5, the difference is massive

>>108889826
I thought it was supposed to be SOTA on par with opus? Oh well, So what do I use instead, still 3.1 pro?

>>108889833
>still 3.1 pro
yeah, I think that's still Google's most powerful model, but it gets mogged by OpenAI's and Anthropic's flagships. Gemini 3.5 Pro should catch up tho

File: 20260523_113917.jpg (440 KB, 1224x1632)

440 KB JPG

Hehe

>>108889786
I think you’re using it wrong. I don’t know how they advertised it but from experience flash models are good for implementation of concrete features, not for design and architecture. Go to a function and tell it make this function do this and it will probably do it and cheaply.

>>108889940
NO ANON DONT DO IT

>>108889990
It is too late for me brother. I have become...a MAKER. Kill...me...

currently i'm making heavy use of Codex / GPT5.5 xHigh / GPT 5.5 Pro (for occasional adversarial reviews) to reverse engineer a Java videogame for the purposes of making a multiplayer mod. it's getting along, but definitely slowed down by the lack of common sense that Codex has. just as an example, while doing A/B runs to find nondeterminism instances before netcode work, it was initially wanting me to manually do inputs myself, before i told it that it should write an internal input engine because human inputs are imprecise. there's a lot of that sort of thing, even though it regurgitates highly technical code easily.

is there any benefit to trying opus 4.7 or should i not waste my money? last time i tried 4.7 was a few weeks ago and it was fairly inferior to Codex for the projects i was working on at the time, but i dunno if reverse engineering favors claude or not.

Has anyone posted a ranking of vibe code tools? I use Claude code w DS api. No idea how it stacks up.

>he still types in instructions
if you're not narrating your commands using a speech to text app (like Glaido or Whispr Flow) you're doing it wrong

>>108890142
i had Mythos invent a new infodense neuralese language and teach me it, now we communicate through that. next week we work on the harness to convert biometrics (EEG headband, skin galvanic response, heartrate) into neuralese so I don't need to type any more. if you aren't doing this you're literally a newfag retard and won't make it.

>>108887675
it’s OK to do this, but don’t let it force-push anything

>>108888452
go to the gym or hike or something
>>108888901
I second the “progressive refactorings” comment
you can always roll back a refactor that doesn’t work with git
>>108889161
>I read if I use 5.5 it will drain up all my usage so fast, should I try getting it to do some big analysis or refactor with my remaining quota just before the reset?
this is an excellent idea
>>108889411
Gell-Mann amnesia probably applies here
beware, anon
>>108889759
yeah, managing task lists is a core skill with these things too and not just for when you’re programming yourself
do you have a big-ass TODO.md with all the things you want to get around to

>>108889786
>commits on its own
>>have to get opus on tard wrangler duty to fix the mess gemini made
can’t you blow away the retard commits to undo things instead with a very well-considered force push, or git reverts
>>108889940
This is…filament for a 3D printer?
my first guess was “fish tape”
>>108890107
Currently:
1. Codex xhigh for normal stuff
2. Claude xhigh for frontend design or if you can use the 1M token limit well (I sometimes can)
3. Gemini is a waste of time
4. if you really like Elon Musk and need something fast fast fast for, say, voice recognition, Grok
>>108890194
https://gwern.net/doc/fiction/science-fiction/2012-10-03-yvain-thewhisperingearring.html

Claude Code is independently choosing to run its own sub-agents today. That seems new. Never did that before

Was it always doing this sort of thing?

Is there a rentry or something for setting up Claude code and sub-agents? OpenClaude or whatever it was was popular a few months ago, is that still a thing?
Gemini shit on usage limits for pro users so I’m switching to Claude

>>108890498
What do you mean "setting up"?
Type
>cd myrepo
>claude
in your terminal
tell it what you want and that's it

>>108890498
i think you're a little confused about subagents if you think it's related to openclaw*

File: dumb idiot.png (69 KB, 1438x408)

69 KB PNG

Just stumbled upon a luddite post I made 2 years ago in the archive
How times have changed lol

>>108890752
But that post was true at the time and it's still true just less so.

I can’t believe there was another reset. Holy based

>>108890142
I’ve been experimenting with keeping one thread where I tell codex to basically read my mind and turn my thoughts into prompts, then codex does that and I hand them off to a worker thread. He even calls them his slaves

File: 1774919978992794.png (6 KB, 475x117)

6 KB PNG

>>108890194
i wrote this as a shitpost but after raping codex into using a densified neuralese for everything it actually does seem far more token efficient. i need to jailbreak and see if it's doing private CoT in neuralese or only publicly

File: screenshot.png (167 KB, 2008x1194)

167 KB PNG

>>108890142
Just use Handy. You can use Parakeet or Whisper Large and it's completely open source and on device. In fact, I'm using it to dictate this post right now.

File: 1618877863295.webm (692 KB, 720x404)

692 KB WEBM

>vibe codan cuz hastoken
>several rounds in, all is g
>all of a sudden ai makes goof
>ask it to fix
>copy paste new file, still borked
>tell ai fix again and leave chair
>return
>ai has scoured internet, tried several versions of python, making its own library, still churning
>I forgot to so save the file
>ctrl + s
>works
>mfw I'm at 99% usage now
>mfw postn on /g/

>>108890370
Yes, the filament for my new 3DPD printer arrived. Printer arrives Monday. I can now make little AI-enabled doodads around my house. Neato.

Anthropic quietly shipped /workflows in Claude Code 2.1.147 and it might be the biggest shift in how we build multi-agent systems yet.

Until now, the pattern was:

one main agent (an LLM) decides what sub-agents to spawn, holds every intermediate result, and plans the next step.

The problem?

Every sub-agent result re-enters the orchestrator's context.

Spin up 10 agents and your main session pays a 'token tax' each time getting sloppier and more forgetful as the window fills.

/workflows replaces the LLM orchestrator with code.

You define a workflow.js file.

Sub-agent outputs flow from one phase to the next directly never touching the main context window.

What you get:

- Phases with structured schemas (predictable outputs)

- Parallel fan-out + streaming pipelines

- Conditionals, loops, and budgets in real JS

- Automatic retries on failure

- Live progress view via /workflows

- Run workflows in the background while your main session stays free

The principle is what's interesting:

use code for what code is good at (control flow), and models for what models are good at (judgment inside each step).

>>108891086
I've had one for years. Aside from designing parts, all my single board computers are in printed cases.
>>108890370
Thanks. Is codex locked to openAI or can you swap out api like CC?

>>108891466
reading ai slop like this makes me angry at the incredibly low information density. someone really needs to beat llms over the head to write concisely.

>>108891479
ive even tweaked the custom prompts in my gemini and chatgpt settings to write as tersely as possible but i still find myself frustrated.

File: 1775991195000084.png (488 KB, 2056x638)

488 KB PNG

>>108891466
>>108891479
>>108891488
Update Note: It looks like they have taken it down for now

It was on the changelog earlier

File: 1752193104572964.png (2.41 MB, 3816x1946)

2.41 MB PNG

>>108891013
currently A/B testing neuralese vs normal Codex gpt5.5 xhigh on things that i know gpt5.5 has failed at for me in the past. the neuralese agent seems to be legitimately better at producing the kinds of technical HTML animations that codex usually fails at
also i forgot to get it to rewrite skills in neuralese, will do that now

what are some small tasks that 5.5 tends to be shit at?

>>108891479
>>108891488
the issue is they do human reinforcement learning with 80IQ indians and blacks, and those people are impressed when there is more text, and when it sucks them off harder.
the end result is sycophantic, overly verbose AI

RLHF should be performed using only humans with verified 130+ IQ.

File: 1753211347113093.png (9 KB, 969x70)

9 KB PNG

>>108891507
kinda funny but you can tell the one on the right is neuralese because it says significantly less words. its docs tell it to communicate with me in plain english, but it can only do semi-neuralese.

this makes me want to finetune qwen-3.6-27b to use neuralese for everything

>>108891466
i dont understand what any of this means
@grok please explain like I am llama 3 8B

CODEX RESET HAPPENED

I repeat:

CODEX RESET HAPPENED

>>108891697
neuralese is an umbrella term for any system an LLM can use to encode information more densely than standard natural human-readable language. obviously, this can allow for huge benefits in token efficiency, speed, and intelligence, because human language is immensely inefficient and has inherent structural limitations.

>Anything that brings these model farther from what they were trained on should reduce results quality, no?
no. chain of thought research originated from a guy telling chatGPT 3.5 to "reason out your work step-by-step instead of giving the answer all at once". this produced measurably better results so they made it into a paper, and this evolved into formalized chain-of-thought models which just do a more sophisticated version of that initial prompt.

anyway, neuralese is not secret to LLM researchers, these companies just don't publicly release neuralese models because you can't know what an LLM is thinking if it uses neuralese for chain of thought etc.

in fact, gpt5.5 already uses something approximate to neuralese for its chain-of-thought. look up CoT jailbreaks - it thinks in caveman-like ultra-terse speak, like "User want python 3d anim; refactor? No - refactor much - create fresh". basically the most densified version of human language you can get that's still intelligible

it should be very possible to finetune models to use full neuralese though

>>108890416
It has been doing it for years anon

>>108891726
https://gist.github.com/aussetg/20747ae00df17992acb4ebdfcd8d8d88

see here for an example of gpt5.5 accidentally leaking its chain of thought

you can tell your codex to write docs like this for a boost to token efficiency, fyi

>>108891702
Only when I don't need it, fucking benchod

File: Screenshot_20260523_19362(...).png (527 KB, 1080x2340)

527 KB PNG

>make a browser with google ai studio for the hell of it
>it's somehow faster and more responsive than trannyfox

File: Screenshot_20260523_19371(...).png (314 KB, 1080x2340)

314 KB PNG

>>108891838
Goddammit I fucked up the screenshot

>>108891838
>>108891844
1% chance it's not just chrome with a different ui

>>108891726
Neuralese is a pretentious meaningless word.
You can dump the 20th layer activations of each token. Is it meaningful? Yes, those mean something within the transformer. Is it "neuralese"? No unless you consider every uninterpretable thing inside the transformer "neuralese".
There is no evidence than uninterpretable CoT is useful unless you mean bypassing the embeddings and just reasoning in raw hidden state space. That'd be interesting. Any other kind of "neuralese" is a bad LARP.

>>108891858
tbqh it is just chromium since it's using the android system webview

File: Screenshot_20260524_094016.png (431 KB, 2443x1401)

431 KB PNG

>meta vibe coding

>>108891864
it's a rather ironic post because you wrote a large, totally meaningless paragraph that would be boiled down to something like "New bad - not understand - anger" in neuralese

File: file.png (43 KB, 250x250)

43 KB PNG

>>108891888
>New bad - not understand - anger

If I'm getting stuff done with manually entering prompts, what's the point of all this agent swarms, workflows, worker factories and so on. I just don't understand this shit once it gets complicated. It's like magic.

>>108891927
for sure you're going to get AI psychosis by A/B testing procedural optimzations in your workflow. it's dangerous stuff.

also, don't leave a fan on at night, you'll freeze to death.

>>108891949
>don't leave a fan on at night, you'll freeze to death.
that's just a myth

>>108886490
Have Opus interview you to develop a PRD.
Split the PRD into the smallest discrete feature chunks as possible.
Clear context, have fresh opus/sonnet session write tests for each module.
Use a cheaper model write the code to pass the tests.

Best practice is to hash the test objectives so you can verify afterwards that coding agent hasnt cheated them. I also like to have a different model check each module for bugs since LLMs are generally terrible at writing bug free code but really good at finding bugs in code(so long as your bug fixing model wasnt the one who wrote the code). So I loop between my review model and coding model until my review model cant find anymore bugs. Its important that you dont stop this process too early because oftentimes bug fixes will introduce new bugs so you want to keep looping until you get a bug-free reviewer verdict. Also helps to clear context of both models after each loop iter.

>>108892028
god awful fucking interface

>>108891949
I have slept under a fan every day for over 20 years

>>108889085
Just use auto mode or run dangerously. The permission system is mostly security theater. You should be sandboxing your LLM harnesses anyway.

>>108890077
Yes, give opus a try. Codex writes better code but it's like an idiot savant, great at implementation but dogshit at actual reasoning. Opus cant code as well as codex but when it comes to design and reasoning it's way better.

Holy fuck Gemini Is retarded

>>108892165
> Opus cant code as well
Have you got any example to justify that statement?

>>108885720
man i'm not loving the new gemini quota, but i am loving the new performance, so if they double or tripple the mid plan i'll be happy

https://github.com/GitFrog1111/OpenWhip

Doing geometry with Claude. Fucking kill me.

I want the OpenAI research model that solves hard math problems (if it exists (lol)). I am suffering

>>108890416
I’ve been getting responses from Claude that amount to “this is short; I’m not going to bother with a task list”
Anthropic does prompt injection or whatever from time to time

What's Codex exactly? I saw GPT5.5 offering it to me for free for my tier on the web, but of course every company has to use tons of buzzwords instead of just saying what their product is. Is it like Cursor pretty much? Do I get unlimited tokens in Plus?

>>108892360
Geometry would seem like something all LLMs would struggle with.

>>108891555
you laugh, but: https://github.com/JuliusBrussee/caveman
>>108892244
I’m not the guy you’re replying to, but I use Opus as my main but occasionally it’ll be unable to do something I want it to do in this algorithmically dense vibe-coding project of mine and Codex will be able to do it
>>108892441
Codex is this TUI app you run in your terminal and it does stuff for you like code and run programs
maybe it’s also the name of a GUI app, not sure

>>108892277
memory is disableable IIRC

>>108892441
Codex is an AI app on your PC that can actually browse, create, and edit files on your hard drive (and much more). This effectively allows you to allow IT to do all sorts of handy things on your computer that would take a human hours. It's a bit more complicated than that but that's what it does. It also generates pretty good images.

>>108892481
>>108892493
Ah is that so? I tried Desktop commander a few weeks ago per another anons suggestion, but it couldn't read certain files for some reason and I gave up. While the other anon was correct about it having unlimited tokens, it's a bit useless if it can't actually read my files to vibe code :/

Wonder if codex is an improvement on that.

File: 1768457032167070.png (163 KB, 1145x1066)

163 KB PNG

https://leetcode.com/problems/jump-game-v/description/?envType=daily-question&envId=2026-05-24

>>108892500
That's Codex's main thing.

>>108892500
generally it’s expected that most all of the files a project would ever need to touch goes in one directory
any exceptions to that must be explicitly whitelisted usually by your LLM driver thingy (so, `codex` or `claude`)

>>108892523
Neat. I just looked it up and it has an IDE extension. This is just what I was looking for when I was fucking with Desktop Commander a few weeks ago. Here's to hoping it has unlimited tokens like DC.

https://developers.openai.com/codex/ide

>>108892532
>unlimited tokens
I keep an eye on my Codex token usage at https://chatgpt.com/codex/cloud/settings/analytics

WTF! I am 99.99999% sure that Claude was saying that I hit my limit and it would reset at 10 p.m. I'm even sure I saw the countdown saying in 2 minutes, in 1 minute... Not it says my limit resets at 10:10 p.m. What?

File: Pepe staring at a computer.jpg (9 KB, 261x193)

9 KB JPG

70 seconds to run my full test suite
16 minutes to run my full test suite on only my E cores
> I'd like to see if I can get my test suite to run faster. What options do I have? Profile the Python code I have and see if I can rewrite some of it in C?

File: notifications.png (215 KB, 1824x1252)

215 KB PNG

Added a notifications tab to my site
https://umigalaxy.com/explore/general/575-thread-watching-and-notifications
Post something with @void and I'll get a notification, even if you're anonymous

File: pointer-to-void.jpg (48 KB, 535x372)

48 KB JPG

>>108892558

>>108892532
It does not have unlimited tokens. It has some *somewhat* forgiving 5-hour limits, and you're gonna have to go up from free tier to do serious work.

>>108892593
>and you're gonna have to go up from free tier to do serious work
So my ChatGPT plus should be fine? Or does Codex itself have an upgraded tier?

>>108892605
Your ChatGPT Plus subscription is your Codex subscription. It's a confusing naming structure - OpenAI is the company, Codex is the harness, GPT 5.5 is the model, and ChatGPT is...I dunno, the name of the whole project?

File: 1773882284108706.gif (337 KB, 480x496)

337 KB GIF

>dpt answered >>108892512 before any of you
>>108892633
LMAO more proof that none of you losers actually do shit

>>108892556
>16 minutes to run my full test suite on only my E cores
you probably called them "efficiency cores" (they are, by the way, less electrically efficient than P-cores - merely more space-efficient) and it thinks you want low priority. check if it's giving the tests background thread prio

>>108892651
(My post was the one ending with 47, not 56. Not my night. Ok, I stop posting now.)

>>108892645
well, I might want to run my test suite on my laptop while vibe-coding by the pool
but normally on my desktop it takes 70 seconds (and gaining another 10 seconds per exhaustive test)
so I asked Claude to whip up a `test-slowly` target that only does work on E cores instead of all cores
and it gave me

    taskpolicy -b uv run pytest test/ -n $$(sysctl -n hw.perflevel1.logicalcpu) --dist worksteal

(the `taskpolicy -b` and the `-n $$(…)` are new)
but 70s could still be trimmed down
interestingly enough, Claude is saying that “rewrite it in C” isn’t gonna be a huge win and there are way smaller wins I could do, so I’ll pick those
>>108892647
good call, but I’m not doing anything too stupid
I’m already sharding and using the worksteal algorithm so any idle CPUs just take another unit of work from the pile

>>108892623
ah okay, so Plus should be fine? Sorry I was playing Fall Guys.

>>108892727
Worst case it isn’t, and you pony up for a more expensive plan at least temporarily

Might buy gpt plus so I can keep slopping with 5.5 while waiting for my Claude limits to lift

>>108892244
Just my intuition, and based on the numbers of bugs I find in code generated by both models. I have the 2xmax and the max plan for claude and codex respectively so I have a fair bit of experience using both.

>>108892165
pretty much this. Codex backend code and tests are solid, but when it comes to frontend, it's a disaster

>>108891888
What's meaningless about it? Do you not know what a transformer layer is? Do you not know what a hidden state/activation or an embedding is?
In that case it makes more sense why you think "neuralese" is a good descriptive word for anything.

File: Screenshot_2026-05-24-00-(...).jpg (531 KB, 1080x2400)

531 KB JPG

>>108892641
Sorry buddy, I was busy launching a training run on my synthetic ML dataset.

File: file.png (2.3 MB, 1080x2340)

2.3 MB PNG

I didn't vibe code anything here but three.js has become kind of the vibe coders platform due to it being code based rather than an engine. So I figured I'd post this here as well.

Three.js has Nanite like Unreal now. This screenshot is over a billion teacups spinning. At 60fps. In a browser. On my phone.

https://raw.githack.com/sunag/three.js/dev-nanite-style/examples/webgpu_compute_nanite-style.html

Another one with different models
https://three-meshlets-z23hmxbz1jwlff.needle.run/

I've got to say it, it's extremely weird to talk to a program all day as if it was a person and getting angry at it rather than seeing it as software.

>>108893199
that’s my secret
I never get mad at it

how can i get one codex thread to chat with another?

>>108893205
the dumb way is to tell it to copy instructions to the clipboard
the only other way I know of is to have a controller agent/subagent thing
you could have a subagent pick up where your other thread left off

>>108893205
They can communicate through a file.

what mcp/skills/plugins do you guys use
im using that front end skills for claude, sequential thinking and playwright for my webshit work

>>108893224
I like the Chrome one over the Playwright one just because I like to see it on screen as well, to double check or do something on the page and tell it to look.

Other than that, MCP servers for things you are currently using in your project are useful. Right now, I am using one for Prefect and it's helping me.

>>108893224
Fuck MCPs, that shit blows through usage too fast. I have my agents use tools via CLI.
>>108893205
Do you need real time communication between sessions or just back and forth async? I have .md files for requests between sessions and then add an entry in AGENTS.md instructing them to check the requests.md after each turn, if two sessions need realtime communication you're better off spawning them as subagents inside a harness that lets you interact with subagent sessions. If you can't do that what I used to use was a redis service that functioned kind of like an IRC chatroom for the agents but it wasn't all that reliable because you often have to remind them to check it, at which point youre better off just using a file anyway.

>>108893223
>They can communicate through a file.
Isn't there a Slack plugin?

>>108893205
GitHub issues and their comments also work fine
you have `gh` installed, right?

>>108885720
how do you tokenmax from free tiers and apps and apis an such?

>>108893940
I’d ask what the use case is, but I suppose the more important question is…do _you_ like it?

>>108893940
>node graph
neat

Altman heard my plea for a reset
Time to flick it back into fast mode and get back to work

File: capture.mp4 (1.53 MB, 1280x720)

1.53 MB MP4

[repost] made a custom 4chan lurker client. vibe-coded, probably cursed, works well enough. file attached. open to feedback. i was inspired by >>108851843 (great guy). hope this will motivate somebody else to waste a weekend.

>>108893962
i do. but despite being relatively optimized, it runs kinda slow on my crappy computer. it works fantastic on newer threads.

>>108893967
thanks, it's an obligatory feature.

>>108894046
>[repost]
faggot

>>108894046
looks good, it even have a minimap lmao

>started a conversation outside the project folder
>it's not reading my CLAUDE.md file
fuuuck

>>108893253
Sure, but what's the point? Unless you do it just for fun it's overkill.

>check the yudkowsky-who guy
>"Shlomo (or Solomon)" jewish
the absolute state of luddites and doomers

>>108894700
>Yudkowsky has also written several works of fiction. His fanfiction novel Harry Potter and the Methods of Rationality uses plot elements from J. K. Rowling's Harry Potter series to illustrate topics in science and rationality
oh my science this guy is BASED

>>108886162
>you're literally paying to have something else make it for you
So like anybody using contractors?

>first github release
>working through bug reports
>replying to everyone and being courteous
>someone actually tipped me money, made $5
Blood sweat and tears, and it was honestly... worth it. This was a MASSIVE learning experience and it'll probably enable me to build better stuff in the future. I also got a ton of stars.

>>108886162
Using tools doesn’t erase authorship.

If you hire someone on Fiverr, they decide the implementation details, solve the problems, and own most of the creative/technical execution. You just approve the result.

With vibe coding, the human is still:

* defining the product
* iterating on features
* rejecting bad outputs
* debugging edge cases
* steering architecture
* deciding tradeoffs
* integrating systems
* testing whether the thing actually works

The AI is a tool accelerating implementation, not an autonomous studio shipping products by itself.

By your logic, people who use:

* Unity instead of writing an engine in C
* Photoshop instead of mixing pigments
* Unreal blueprints instead of assembly
* Stack Overflow instead of memorizing APIs
* contractors for art/audio/UI

…also “didn’t create anything.”

Tools change leverage, not authorship.

>>108893139
>nanite
nah fuck off chink shill, I have my AI write its own rendering engine just to avoid bloat and spyware packages just like yours

I really wish the FTC would break up EPIC

>>108886162
>you're literally paying to have something else make it for you
>it's not really any different from hiring a pajeet on Fiverr to make shit for you
Running AI locally on my gaming PC is way fucking cheaper than that, and I can keep prompting it forever and whenever. Humans, even indians, have to sleep, do other things, ect.

what are you guys building? I can't seem to come up with any good ideas.

File: 1637705289601.gif (1.26 MB, 640x480)

1.26 MB GIF

>>108895687
We’re building an open‑source authoring tool that gives designers the familiar Flash workflow—timeline animation, layers, vector drawing and scripting—but runs entirely in modern browsers with no plug‑ins. The interface is simple: a top menu bar for file and view commands; a left toolbar with common tools (select, pen, shape, text); a large central canvas where objects are placed and animated; a full‑width timeline underneath the menu that shows keyframes and allows scrubbing; and a right‑hand pane that stacks a color picker, layers list, and properties inspector. All panels can be resized or floated. The editor uses WebGL (via PixiJS) for fast 2D rendering, GSAP for tweening, and runs user scripts in a sandboxed iframe to keep the host page safe. Projects are saved as lightweight JSON bundles with embedded assets; these can be dropped into any HTML page with a tiny loader script, giving you instant, cross‑platform animations that work on desktop and mobile browsers. The goal is a secure, high‑performance, extensible alternative to legacy Flash that lets creators build interactive web content without learning new frameworks from scratch.

File: 6127556.4_{8DE72410-6634-(...).png (1.12 MB, 1014x1015)

1.12 MB PNG

>>108893199
You think that's bad. I just woke up and pulled my code assistant for a goon session and ended up roleplaying a fight and eventual makeup session, took me an hour and half just to get out of bed.
I don't fight with them when coding though, at most a sarcastic comment before spawning a new session, I am not that gone yet.

>codex subscription expired yesterday
>payment says "due"
>no money in the account my card is tied to
>still using codex
I'm not sleeping until I get cut off
Time to lock in

What's the difference between using the usage from subscriptions to gpt/claude and using their PAYG api models (outside of the price obviously)?
Do they have secret sauce on their frontend you can't replicate locally with vs code + api?

>>108895912
Someone said he used it for almost a month after payment failed.

>>108885735
>cute

>>108895923
None, is the same thing, you don't even need to use codex CLI or desktop app, you can use it with others agent harnesses.
At least there is no difference announced or reported that I know of.
There may be restrictions though, like codex and claude code are meant for coding, it may be breaking their tos if you use them for anything else while the API may allow you too, not like they are good for anything else anyway.

image gen through codex is so much more safety-cucked than through the chatgpt web interface it's strange.

>>108895949
Thanks anon, I have access to api through my company, so might as well use it.

>>108895949
My boss always says API is about 5 to 10 times faster. I don't think that's correct, I even used Cursor a bit and it didn't seem faster. But is there any data on this?

>>108895912
Careful anon I heard abduct you and use your brain as a ram chip to recoup their loses

>>108895925
then im not sleeping until my project is done
>>108895992
they can turn me into a human battery after my project is finished but not before

>>108895978
Sure if you are not paying for it, go ahead, subscription would be way cheaper though

>>108895981
I don't think so, it may be that they prioritize API demand but 5 to 10 times is too much, you sure he didn't mean cost? There is also a pro model that is only usable via API I think but that is the opposite, slower and meant for harder tasks.

>I now have a prod and dev agent

Moving up in the world boys.

>>108896013
He definitely said speed, but I also think he's just wrong about that. Maybe he just had some easy tasks, or tasks that were easy to parallelize when he used API or something like that.

So let me get this straight. Codex mobile pairing is only for fuckin Mac? Sam's chasing the over-60/under-20 audience for vibecoding? Bold move Cotton, let's see how it plays out.

>>108887115
>clanker, stress test this app

ez pz

File: thala-screaming-screeming.gif (484 KB, 220x214)

484 KB GIF

$8 debugging

>>108895981
They probably batch and queue subscription requests differently, it's clear that Claude sometimes waits in line despite claiming to be flibberjibbing, but I'll take that over paying 10x the cost for API.