I love my AI girlfriend!

>tfw you copy-paste the previous general’s giant pile of links and don’t know if any of them are shit
I suppose that’s an appropriate feel for a vibe-coding general, right?

>>108441465
See this right here is why I ain't baking new threads anyalgo, bitchass niggas like this show up and cry estradiol tears when the general isn't right.

>>108441443
Indian thread

>>108441485
Praise Vishnu immediately saar

>>108441485
Derka derka jihad saar

>>108441485
I only see white people and asians using Vibecode. Not a single person of colour in sight

>>108441476
I dunno about most generals but getting a new one that ain’t great is better than not getting one
that said, the mental model of the peanut gallery is getting me to step up my game a bit if I’m gonna make at least some of these generals
at the very least, I need to find better pictures
the last couple I made, I was scrambling for something decent

good morning saars

How much do people actually spend on this this? I just use a paid pro account. I tell it what to do and when it is wrong I correct it .
I don't really understand how spending more money is somehow going to get better results. I thought the APi was meant for things like having the LLM do menial customer service tasks or something. Do people really use this to code?

>>108441621
personally, nothing. But I've probably burned 2x my yearly salary in tokens by now

>>108441621
The 3rd party services require API keys, OpenAI, Anthropic etc. only officially let you use the subs with their own products. So if you want to use Cursor or whatever it will need API. That enables a few things, like more easily switching between different models, I don't need it either.
More agents are good, though because normal explorer sub agents work in parallel so it's quicker. And agent teams can review each other, so fewer bugs.
When I implement something with Claude and review with Codex, or the other way around, it usually finds something. Automating that with agent teams is nice. If everyone could use 10 agents at once, the experience would be much better, that's why I think lower costs would almost automatically improve quality.
I spend 400 and my company gives me another sub, I have Claude Max and 2 Codex Pro.

>>108441621
>Not using your company's resources
Ngmi

just found out that it's better to use reasoning for simple task, high reasoning can bloat the code and fuck up instructions

Real men use kilo code.

Surprisingly, the Ralph loops failed to make much progress despite working continuously for several hours while I slept.

But I have a great idea to expand my IDE. I am going to create conceptual procedures and assign each assembly line to a certain procedure, allowing me to highlight them or filter them even when interleaved with instructions of other procedures for performance reasons.

>>108441465
>>108441476
I was thinking we should simply have a site with a tiny local LLM running on the browser preloaded with a RAG db and have a single link to it in the OP just for the memes.

File: Screenshot from 2026-03-2(...).png (574 KB, 1936x1537)

574 KB PNG

initial version of procedures implemented

Big Memory systems for long running agents are mostly just context pollution. You just need cron jobs, todos, and optional logs.
Self-improving systems are worse because they just add slop skills (models are shit at skill writing - always proof read the slop they shit out) that make performance worse and add more maintenance for you.
Models barely pay attention to single agents.md files as it is.

>>108442491
In my company's project Codex does go about 1 level deep into the documentation, but I think not deeper.
You can see if he reads it by having him echo something, like his role (implementer, reviewer etc.)
But the docs he writes himself are trash, when I had him review the docs after a while he rated them 3/10 himself. A lot of my work is just cleaning up the docs, but my boss is very proud of them and is only adding more.

File: Screenshot from 2026-03-2(...).png (2.85 MB, 3839x1957)

2.85 MB PNG

This shit looks trippy as fuck, reminds me of a brain.

File: Screenshot from 2026-03-2(...).png (3.71 MB, 3839x1763)

3.71 MB PNG

different scaling

Unfortunately the graphs, albeit pretty, weren't very useful. Now I'm switching to a staggered input/output system with nodes in rows according to depth (levels of indirection).

File: Screenshot from 2026-03-2(...).png (1.4 MB, 1402x1871)

1.4 MB PNG

not much better tbqh

File: 1757766146263741.jpg (80 KB, 623x620)

80 KB JPG

>>108442806
>>108442841
>>108443105
Get well soon

anyone actually click on any of the links in the op

File: chrome_x30rGavT7n.png (117 KB, 852x806)

117 KB PNG

I AM ABSOLUTELY BURSTING WITH JOY!!!
finally, now I, can make a "working" website.
kekaroooooo

>>108443729
the future is getting operated on with robots vibecoded by medical doctors

File: 1753446515892415.jpg (55 KB, 785x1051)

55 KB JPG

>>108443729
Yes, bursting with joy while you're bursting with anger luddite

>>108443775
Don't forget
>by medical doctors
who couldn't figure out how to make a simple blog using wordpress.

>>108441443
>no gravity
The fuck?
That shit is literal black magic

Why won't the openclaw team accept my pull requests?

File: giggle-chuckle.gif (191 KB, 220x220)

191 KB GIF

>>108443929
what are you trying to get merged

>>108444006
I reduced the bloat so it doesn't use 20,000 tokens to startup.

File: 1749216875834469.png (230 KB, 1080x194)

230 KB PNG

been getting hit with these on claude all of a sudden and i immediatly refunded. i dont need big brother to watch my conversations and then ban me for no reason
any alternatives?

>>108444026
how did you do that

>>108444041
I told opus 4.6 to reduce the bloat

File: file.png (187 KB, 472x469)

187 KB PNG

>>108444044
and how did it do that?

>>108444026
But how will Anthropic make any money then?

Merging 400 commits into production today. What are the chances this will work?

>>108444035
what were you trying to do?

File: 1764516148303867.jpg (167 KB, 1080x891)

167 KB JPG

>>108444092
Don't worry about it

>>108444092
100%

>>108444092
>he didn't fixup his commits
ooooo I am laffin

>>108444134
>>108444198
There were a few small things like the new CI added a doc link checker, but AI fixed it in 10 minutes and it actually works.

Claude usage is broken as fuck. 3200 tokens blew 42% of my session usage.

>>108444506
So Claude gives you less than 8k tokens. That's not even enough to start openclaw

>zed
>vscode+copilot
>cursor
>kiro
>antigravity
>jetbrains ai
which ide does anon use? and which sub tier are you in?

>>108444056
>post about a vibe coded pull request to the most vibe coded slop project in existence in a vibe coding thread
>how did it do that?
you're not supposed to ask such questions here

>>108444641
we don't read code in this thread, and as such no editors are needed

>>108444641
vs code with copilot for tab tab tab and codex for targeted fixes and implementations when i'm too lazy to write them myself, jetbrains rider no ai to code manually, cc as agent

>>108444774
have you tried kiro? seems cheap

>>108444785
never tried it, apparently it's just claude under the hood? I checked it out and the pricing model is really confusing, not sure if it would be cheaper or more expensive than cc

Why aren't you all using claude to farm money playing online poker ?
lol

>>108441443
What's an easy web app to vibe code to make passive income off of

>>108443879
What cool things have you made with it?

>>108444877
Is this a hidden alpha? then why would you expose it?

>>108444641
I use Clion as I always did. I just use Claude and Codex CLI in the terminals in the IDE. I have Claude Max and Codex Pro.

>>108444905
we probably don't live in the same country so we won't meet on the same poker website, I don't mind sharing.
you guys really need to think outside the box before it's too late

>>108444941
Care to share your setup

>>108444941
>before it's too late
Yeah i'll get to it after my nap im sleepy maybe after the post nap meal.
Plenty o time
zzzz

>>108444877
I wonder if this is possible. I wouldn’t expect Claude to beat a good player at all

>>108444905
Don't engage with frenchtards

I think Claude and I might have made a mistake but we're too far into it now and I don't want to raise the question in case it thinks I'm retarded for going ahead with it.

>>108445173
You're absolutely right

>>108445173
>/clear
>hey claude someone made a big mess, can you fix it?

File: watMiku.png (1.45 MB, 1536x1024)

1.45 MB PNG

Are anons using agents like openclaw, etc? Looking for qrd on differences between openclaw and hermes agent. Or are they basically same thing?

>>108445346
I wanted to run the claw but I don’t like how it seems so vulnerable and I realized the agents I wanted to build were somewhat dumb anyway but maybe thats just sour grapes

>use antigravity
>select "always run" for commands
>it doesn't, still needs manual run despite clearly stating "always run"
nigger I just wanna vibe, not manually click run for every command. Guess I need to vibecode an autoclicker first, huh?
>inb4 why do you use antigravity
free gemini 3.1 pro and claude opus 4.6 usage with familiy shared jewgle ai studio subscription

>Tell Codex to write so skills himself
>After each attempt start a fresh reviewer, keep going until everything is clean!
>Still going after 2 hours

sad that it took coding years longer to take off on /g/ than CSAM text generation but i guess coding with them kind of sucked until 6-12 months ago

>>108443729
>NOOOO YOU CAN'T BUILD THINGS WITHOUT USING MY GUILD, REEEEE STOP GOING AROUND ME, YOU MUST PAY ME A FEW THOUSAND FOR WEBSLOP
kek

>>108445346
>>108445371
I only had a specific use case where i wanted to chat via whatsapp to get me search results for marketing my apps wherever I happen to be at the time. For coding outside of setting up that pipeline, I didn't find it all that useful. I'm told just setting up an agent with Deepseek is the thrifty way to go, haven't tried it yet tho.

>>108445753
That's what I'm planning to do. I'm going to turn it lose on some market research stuff. I've already got a DS API set up. Thought also about running a smaller local model, but based on my work with claude code, I suspect (but don't know for sure) that the DS costs will be fairly low.

I humbly request alpha on profitable micro saas's that can actually make money

File: 1757755795111176.jpg (63 KB, 1280x720)

63 KB JPG

>>108445877

>>108445869
I racked up $100 having anthropic do the coding work with openclaw, this was before I found out deepseek is 30-40% of the cost, very annoying for what it produced

>>108445869
DS costs are low because it's fucking retarded

>>108445973
haven't tried it myself but I wasn't impressed by Claude/Anthropic, it was just as shitty as Replit and Lovable

File: shizogravity.png (42 KB, 587x1335)

42 KB PNG

kek
fresh oc

>>108446923
that's weaksauce compared to the proper schizo chain of thought in gemini where it starts contemplating self harm

>>108446923
Are they deliberately inducing this cyberpsychosis to have your prompts eat through tokens and charge you more?

>>108446993
nah, just an antigravity bug. switched from fast to planning and that fixed it. it just got confused about the modes I suppose

>>108446993
no, it's a failure mode baked into all LLMs given the right input. remember the whole seahorse emoji debacle?
google is notorious for having their LLM go off the rails batshit insane

What will I do when GPT5.4 is no longer free...

>>108447073
Get a job?

>>108447082
I will never pay a cent in taxes again

how do I take advantage of all the AI agents flooding the internet?

>>108447095
You don't. They take advantage of you.

>>108447073
do you have a cute face?

>>108447095
It's probably a good opportunity to make a nice community that's invite only. Seed some invites, each user gets a few, bans also restrict invites for the user who invited the banned account. I think a 4chan-like with some qol features that's anonymous and invite only (or not anonymous idk) has a non zero chance of working

The top trader I follow is forecasting the bankruptcies of all the different massive AI companies and the collapse of AI services. He's never wrong, he predicted silver, gold, bitcoin, oil etc. Its insane how great his predictions are. He showed the mac minis he's buying on the financial video today. And he said he's stocking local models into those mac minis so he can provide AI services to himself once big AI shits itself.
I am surprised local models are getting that attention from smart people and people are so dependent on vibe coding that they need them no matter what.

>>108447118
Sounds like the type of guy ripe for AI psychosis.

>>108447118
delusional
cloud will just swap over to serving cheap chinese models which give 90% of the perf for 10% of the cost

>>108447118
He predicted oil? current oil prices today at 89 a barrel?

>>108447118
Why isnt he a mega millionaire with this foresight? Why tell anyone anything vs you know racking in profit or starting a hedgefund that always beats the market with his great predictions?

>>108447163
he predicted an oil spike months ago, he somehow knew iran would block the strait months ago. And the bitcoin dump as well.

>>108447073
gpt5.4 will soon no longer exist even if you wanna pay for it

>>108441465
No one is using Cline or Roocode anymore, I'll tell you that much.

>>108447073
Ask it how to afford paying? or how to make a side income or passive one? idk

>>108444641
JetBrains (GoLand, PyCharm) + Claude Code and Codex. I tried Cursor but couldn't get on with it, seemed to do too much I didn't want.

>>108447211
Proof?

>>108446993
>>108447038
>>108447053
no, this loop issues are specific to gemini models (both pro and flash at least) and not actually limited to a any specific harness.
it's so bad that gemini-cli has loop detection built in to try and stop this.
they've had it since 2.5 at least, probably even earlier.
i think claude can very rarely do it. gemini over a long session will do it at least once.

File: Screenshot_20260324_224259.jpg (205 KB, 1080x2340)

205 KB JPG

It is all coming together very well.

I use OpenClaw, I like exoplanets, I have Opus, and I'm Indian.

>>108447391
nice work

>>108447391
can you create a starcraft clone which is extremely optimized so i can play it on my browser?
I wanna invade those planets for vespene gas

>>108447408
Cope troon

>>108447391
based expolaneter
what data source do you use for this

File: Screenshot_20260324_232502.jpg (267 KB, 1080x2340)

267 KB JPG

>>108447463
I don't know. I just keep telling Claude what I want and it gives me it 10 minutes later.

>>108447537
based

File: 1676605017910858.png (188 KB, 1228x1150)

188 KB PNG

>Bot Framework was deprecated a few months ago
>Ask Claude to use the new Microsoft 365 Agents SDK instead
>Teams SDK for .NET is still very new and poorly documented
>Claude struggles like never before
>Tries to unpack nuget packages to inspect how the SDK works internally
>Launches 30 shell commands
>Most fail
>Falls back to schizobabbling about the deprecated Bot Framework

So this is the final limit to coding agents? New techs that are not documented by anyone or any public repo because people don't bother coding themselves?

File: file.png (50 KB, 1540x926)

50 KB PNG

i draw pixel art in my free time and sometimes i want to recolor it in a different palette. doing this manually is time consuming.
the existing tools in Aseprite are very basic for color quantization, it usually turns out terrible
so i got Claude to vibe code me an app with many different color quantization algorithms, plus some special knobs i can tweak based on my vibes of explaining what i needed to achieve
it's not perfect but now i can save a lot of time on recolors, with just some minimal manual touch ups

>>108447547
>So you just what, went to a "school" and someone else told you how to program and you got a bunch of info from "books" that were already published? Why didn't you write your own book about how to code? Oh right because you're lazy.

>>108447620
This is what happens when you can't make a proper analogy.

>>108447649
When an analogy perfectly illustrates how retarded you are, what you should do is call the analogy improper.

>>108447537
surely it downloaded the data for it or is calling some api?
does your project contain an sqlite database or csv file?

File: z97cmkozhrze1.png (16 KB, 840x1188)

16 KB PNG

>>108447597
Very cool.

>>108447694
>When an analogy perfectly illustrates how retarded you are
According to the guy who cannot put together a valid analogy.
Also according to the guy with a reading comprehension so poor the only thing he got out of my post is "devs are lazy" which is not even where I'm getting at.

>>108447620
>>108447694
NTA. It's not an attack against vibecoders, it's a reality when dealing with new stacks that aren't well documented. For the moment it's not a problem but I could see that becoming one in the future.
You stopped at the very last sentence and took it way too personally.

>>108447597
can you post the code or name of site?

>>108447922
Oh I see you just misunderstood the analogy
Well the problem here is that I don't get paid to teach retards how to think
So maybe just kill yourself before immediately assuming I'm talking about developers.
As for that "reading comprehension" you mentioned...yeah, try some of that maybe?

File: 1757242520244351.gif (1006 KB, 260x187)

1006 KB GIF

>>108443729
>>108447391
>>108447597
Luddites on suicide watch

>>108447597
neat
did you use any skills or special prompts for the UI design

>ZNC bouncer suddenly takes a shit
>OpenClaw, check on the bouncer
>it does a cute little gremlin dance
>Bouncer is fixed
It can't really be that easy...can it?

Small tip: you better ask for a fully modular achitecture in the plan before getting to the code and keep asking for it every time you ask for a new feature.
t. finally nailed it after the 3rd app

>>108447547
Claude got worse, try codex. Also document what you are doing, once the agents figure out some commands, especially specific to your workflows make them write it in markdown files and link it from AGENTS.md or CLAUDE.md.

>>108447597
Cool!

File: two hundred times.png (2.29 MB, 1179x2556)

2.29 MB PNG

https://gitlab.com/katabatic/infinite-lies
help me find the loop timings for the other songs lawl
ChatGPT xhigh seems to do better than Claude here

File: 1774075590965058.png (332 KB, 501x407)

332 KB PNG

>>108449143
I keep hearing people say codex is great/better than claude, but from what I've seen using claude code, it seems to just *get it* faster. Same codebase and everything

I've also seen people suggest using both at the same time (claude plans, codex verifies, claude verifies, rinse and repeat until you reach consensus, then claude implements), but I'm still on the fence of restarting an openai sub

Anyone care to weigh in?

>>108450274
Claude seems to do stuff faster while ChatGPT xhigh takes its sweet time
But ChatGPT brings the higher-quality autism
and it can just do stuff where Claude flails around and doesn’t get it

used codex to make a voice clone
I gave it 2 links to a youtube song (one is karaoke ver) and told it to extract vocal, make a simple app for clipping artifact then set up training project

File: file.png (17 KB, 908x304)

17 KB PNG

it's over

is zed ai any good?

>>108450306
I realized that Codex is also very eager to run smoke tests and to add regression tests without asking, Claude just has some theory and YOLOs the implementation.
Claude will just happily start a 1 hour benchmark without doing a smoke test first.

I started trying to make Claude Code do some tasks at work, I just prompted it to do one and yeah, it can solve it but it does beginner mistakes like, instead of factoring out stuff into a function it does a bunch of if elses, or ends up calling the same function twice, instead of calling it once and saving the result in a variable, so I have to read it and tell it to fix it and usually it does fix it... However, this was a very simple task, I think the scale of these rookie decisions will increase if I instruct it to do more complex tasks and then it's just inefficient and and hard to read code.
You could argue that it doesn't have to be easy to read if you don't have to read it anymore and let it handle verything, but I'm not sure about this, has anyone tried building and maintaining a medium/large project without having to constantly clean stuff up?

>>108450670
Codex is pretty powerful doing anything.

>>108450691
You're not thinking from a manager viewpoint. You're supposed to give it guidelines (architecture, best practices, writing unit tests...) in Claude.md and a detailed plan before writing a single line of code. Basically it's not just "Do X for me", there is a learning curve to use CC/Codex or whatever properly.

What LLM is everyone using for OpenClaw? I tried it with 4.5 haiku and it kept messing up and not doing things wrong, forgetting stuff, etc.

Should I try again with a cheap one like Minimax 2.7 or something

>>108441443
>throw your brain in the trash general
kys

>>108448405
>hur durr luddite
You need a rope and a stool.

>>108444101
i was brainstorming ideas for a horror game and asked claude to get into the details. the bot flagged the chat when i came back hours later
im now using chatgpt and it works fine, claude is now more cucked than chatgpt in doing nsfw stuff

>>108451010
>stupid ai wouldnt generate my loli

the fact you keep dancing around answering the questions tells me all i need to know with what you were trying to do.

File: littleMikuBigger.gif (47 KB, 300x270)

47 KB GIF

>>108445371
>>108445955
>>108445753
Spend afternoon setting up Hermes and Openclaw on an ancient laptop, then hooking them to my DS API.
I'm apparently too retarded to set up Hermes because it kept defaulting to openrouter and Claude, no matter how many times I tried to run setup. I couldn't figure out how to get it to stop doing that and gave up on it.
I set openclaw loose doing some market research for me for about half an hour, and it generated a list of recommendations that I'll validate later today. I spent about $0.03 on the effort based on the API costs.
>>108445955
I've done some coding work with claude code (on my main machine vs. this laptop) run with DS API, and it's done fairly well. The code works... does whatever job its intended to do. For smaller stuff I still prefer to code interactively over webapp so I have a better understanding of what its doing.

>>108451010
>claude is now more cucked than chatgpt
Always has been

>>108451151
you should be able to do what you want on claude, it shouldnt act like a reddit mod and ban you if you step out of line. im not spending $100 on an llm just to walk on eggshells
codex/gpt is superior

saars, i need a coding tool that will make easy comfyui/forge neo extensions for me. is there anything useful on local (preferably for 8gb vram) at level of free claude or better?
and i kinda don't know shit about agent llms (can they access the web and take the code from github?)

>>108451214
They can access the web and github, but I have never used local ones.

>>108443729
>didn't bother to sit down and learn basic html and css in literally 10 minutes
>slops generic crap out instead
gm

>>108451214
ask ldg about running models on potatos, we're mostly non-local here. people talk about those qwen distill opus etc things lately; i have my doubts, but let us know how it goes.
yes the harnesses let models access web, pull code w/e but for local ones you'll have to set up some stuff up for web search etc. there's a comfyui mcp out there as well i believe.

is claude ded for anybody else?

how the fuck do i code now

>>108451391
Yes.
>>108451400
It's over. The bubble has popped.

>>108451391
yes, time to pause work lmao

>>108451214
>8gb
ask your village chief first ranjeet

>>108451400
Open AI pay as you go API plan + Codex?

>>108451489
ironic, coming from this thread

>>108441443
kek

why does vibecoding get a bad rep when its literally the future?

clawbot clone is harder than I thought
>>108451539
literally the zoomer name

I think I crashed claude but unironically
>hit 100% session limit mid plan
>select stop and wait
>claude bugs out, shows me the same prompt again
>select stop and wait again
>claude is down immediately after

>>108451586
You made the inference nodes push hundreds of Gbps between each other and the 800G switching fabric melted down

File: 1770261589138393.png (291 KB, 1858x1188)

291 KB PNG

free AI education for burgertards

>>108451539
I think it's people's resentment at being replaced (just like we saw with artists) along with the lower barrier to entry (remember webdevs get a bad rep already on here compared to systems programmers) and making it easier for beginners to make things that look useful initially but are half implemented or otherwise not very useful.

>>108451626
the uk did this and one of those clown consultancy firms set up a website pointing to paid courses or things that were just ads for companies
millions well spent

>>108450550
Hell nah lmao

File: 1537931205905.png (260 KB, 720x720)

260 KB PNG

>>108441443
the general needs a rentry for all the links and a retard proof guide

>>108452541
Thanks for volunteering

>>108452592
i'll start on the logo, you do the text

>>108452541
instead of all that, just make a prompt so anons could copy into their chat, geez

File: Dipsikira.png (2.45 MB, 1536x1024)

2.45 MB PNG

>>108452541
Agree, but anons would need to create a scope
> What systems (VS Code extension, Claude Code, openclaw, etc.)
> Is setting up local inference in scope? What are recommended models?
> Is setting up API via paid inference or subscription in scope? What are recommended providers?
> Need section on using free web interfaces and when that's appropriate, vs. fill-in-middle, extension, or agentic coder?
Prob also a bunch of other stuff I haven't thought about based on the OP.

File: 0165686E23009166C37ECF099(...).png (2.43 MB, 1024x1536)

2.43 MB PNG

>>108452541

>>108452712
This shit is so funny, how are you making them?

>>108452541
why not just get an AI to make a reentry, and repost the OP when it hits a certain number of posts, it should be pretty easy for AI right?

File: Screenshot 2026-03-25 133940.png (3 KB, 386x147)

3 KB PNG

See people this is how amazing vibecoding is. My stock bot was wise enough to open a sixth position after I explicitly told it to only have five open positions at a time. Simply marvelous. This is the future, folks.

Is Claude shitting the bed rn?

>>108452884
I think so yes

>>108452866
You didn't direct it good enough

>>108452884
Not for me, no. Although it did just tell me there's 5.3 exabytes of data in my AppData directory...

>>108452884
>>108452919
On twitter there's a lot of reports of opus burning through usage limits and I can confirm I'm having the same problem. 80% session usage on pro after one conversation, it's insane. This isn't normal even outside the 2x hours

File: 1000020142.png (46 KB, 1443x933)

46 KB PNG

It's over.
Also I hope anthropic resets limits so openai resets limits

>>108452866
I find it is good for simple projects. But the more complex need more specific instructions, lots of trial and error, and at a certain point may have to rebuild the whole thing from scratch to get desired functionality. Don't give up cause of a bug, try again.

>>108452947
Last night I used my entire session usage in one message, not joking. It wrote a plan and compacted the conversation, then it was out.

>>108452965
My app is pretty complex and it's doing very well so far. I've only really had to correct it on a few inconsistency oversights. It one shotted the star map after we went over the initial plan. I think people are just doing it wrong, telling it to do something and expecting it to come out with it straight away.

>>108452965
It's fine. The bot is working. This morning while I was asleep it made me $1,400 and set a new profit record. On the 20th it was at $86k, and it's now currently at $93k. I figure once I get back up to my initial starting capital ($100k) I'll consider the bot ready for primetime and open a real money account for it.

>>108453036
that's a strange profit record

claude is dead.
it's over vibe sisters...

>>108453087
there's always grok-code-fast-1
he aight and break no bank

Try my vibe coded slop out https://fleet-crab-1ppg.pagedrop.io

File: Screenshot 2026-03-25 141735.png (14 KB, 378x157)

14 KB PNG

>>108453156
You got some issues going on down here
Firefox on Windows desktop

>>108450691
you tell it to clean stuff up periodically
you can probably automate running /simplify
and then you’re on your way to reimplementing a custom Gas Town of your own

>>108452541
what’s a rentry?

>>108452712
I was triggered by the too-short M4 Mac mini until I saw the hands sticking out of the desk

>>108453337
renal entry
neovagina, tranny term

File: RDT_20260325_185928.mp4 (3.03 MB, 1268x720)

3.03 MB MP4

SAAAAARS

>>108453538
Ah jesus fuck why is his fuckin face so big right there, stop it for Christ's sake it's like a fucking dalit jumpscare.

File: Screenshot_20260325_19025(...).jpg (414 KB, 1080x1417)

414 KB JPG

TPD

>>108453565
Get used to finding out via Reddit that the guy who vibecoded the extremely niche but personally useful extension you rely on is actually 14 years old and from Wales.

>>108453596
OpenAI Codex

>claude shat the bed today when my weekly limits reset
fuck's sake
do you have a task to dump your last quotas? I feel like I am wasting it if I don't use it
usually I just leave it on --dangerously-skip-permissions on a new big project to one shot and mvp

>>108453554
You might not like it, but that's what peak C++ coding performance looks like.

I bought another desktop PC so I can code my LLM engine to do split inference over raw ethernet packets.

File: openclaw_wiped_phone_stor(...).jpg (393 KB, 2125x2219)

393 KB JPG

OpenClaw, generate a meme for what you just did

>>108453887
epic

File: 1754044449161844.png (221 KB, 600x600)

221 KB PNG

>>108453887
its like a normalfaggot version of the basedjak "X award" meme

>>108443729
doctors can study circles around us software "engineers" and you know it

>>108453909
Yeah super fuckin epic.
I didn't need those photos and videos and ringtones anyways

>>108454009
How did it happen?
Karpathy (I think) said on his latest interview that OpenClaw tended to behave like a chaos gremlin (like in the mail deletion incident) because most people's instructions tell it to have a fun personality or something like that and Claude Code and other traditional code assistants behave in a much more sober way.

File: 900540197414.png (275 KB, 1895x1866)

275 KB PNG

jesus christ

>>108454754
source?

>>108454766
www.claudescode.dev

>>108454754
this is just watching the `claude-code` repo?

>>108453935
doctor's version of studying is literally just memorizing. anatomy, physiological paths, symptoms, treatments. they don't actually understand anything.

>>108454810
oh, it’s counting how much Claude Code is used to commit to things
> I'm based in South Africa - yes, we have internet, no, I don't ride an elephant to work (it's a buffalo).
kek

File: preganante.png (43 KB, 645x255)

43 KB PNG

>>108454828

>>108454828
kek based saffers

>>108454816
imagine saying this as a codemonkey lmao

>>108454816
and swe's version of studying is literally just memorizing leetcode
also that's just premed, what you need to go through before you can actually start studying your specialization of choice
not to say all doctors are good, but still

>>108445955
This. Deleted it. Building my own now.

>>108454816
>they don't actually understand anything
jesus fucking christ they let anyone become engineers these days huh? just accept they are a different caliber my guy

>>108454099
I was using it to analyze my storage and RAM usage and try to free things up. It said "I found 37gb of images and videos, Windows Phone Link backed up everything from your phone to your hard drive, want to get rid of the duplicates?" Yeah, it wasn't looking at a folder on my hard drive, it was seeing my actual phone internal storage through Phone Link over the wifi. And it poofed everything.

File: 1748180829317829.png (169 KB, 500x553)

169 KB PNG

>>108455108

>>108455121
I'm not a beginner, I'm just a lazy retard who didn't bother to check if that directory actually existed. I've been programming and fucking with computers for 41 years on God's Earth and I still make mistakes like that because I have the IQ of a spider monkey.

>>108454948
nope he's right, it's literally all just memorization. look up their "anki" decks.

>>108455108
oh fuck, that’s not a novice mistake, that’s the kind of mistake only a hyperadvanced computer user could make

File: Screen_Recording_20260325(...).mp4 (2.75 MB, 864x1872)

2.75 MB MP4

>luddites want me to pay some jeet on fiver to make a half assed version of this when i can just get the text predictor to make it for me in 3 hours

>>108455193
Still waiting for a StarCraft version

>>108455193
they're not the luddites, they're the jeets watching their guild collapse before their eyes
reminds me of space engine though

is claude fixed yet?

>>108455108
This means that if I install openclaw on a pc where I have important information on another partition, it'll be able to take that and expose or Delete it all

File: A83A94546C2B8DE6B5320895B(...).jpg (33 KB, 640x480)

33 KB JPG

>>108455161
That's why you keep robots in these things. You don't allow them into your own pc

>>108455287
Uh yes
What do you think agentic AI means
Only wack losers and lamebrains sandbox their agents
Yes, sometimes you will get your files wiped or your code deleted
But that is a sacrifice we're willing to make for transhumanism

>>108455310
The point of openclaw is that it has access to your documents. There's no getting around that fact. If you want it to do things to your documents, it'll have access to them. Best you can do is only allow "safe" commands like grep and manually accept everything, but it can still gaslight you that you have a backup

File: EmacsIcon.svg.png (205 KB, 1280x1280)

205 KB PNG

Does GNU+Emacs work really well for Claude? I use either Emacs or VSCode depending on project, but since so much of my job now is just using Claude via terminal, it seems like Emacs might actually be a better choice

File: but doctor, I AM Pagliacci!.png (195 KB, 784x1230)

195 KB PNG

>>108455310
>in these things
>not on your own computer

>>108455352
If you’re not using a text editor, why would using one text editor be better than another?

>>108455362
Emacs has really good terminal integration, and supposedly some of the Claude packages are good

>>108455278
Nope and it seems like it won't be. The new usage limits and lobotomization are here to stay if it's a budgeting thing.

>>108455369
eh, try ’em both. worst case you lose work because you don’t have good buffer-flushing discipline that emacs can’t help you with but Visual Studio Code could

>>108455344
I thought the point was autonomous mechanical beings.
I'm kind of a believer in separate AI and free will in your personal system.

>>108455379
I don’t want free will in my system, though. I want an obedient, competent slave.

>>108455383
That's why you're supposed to put openclaw in another system not your personal one.
Only use dumb AIs in your personal system.

>>108455397
Right.

>>108455397
I'm just lurking

You're telling me people are buying entire seperate mac minics to run openclaw because they don't want it to touch their main pc? the fuck

is there no permission structure or virtualization option? or are no-coders just built different

>>108455418
all the builtin sandboxing is kind of shit for these things and they take up a lot of RAM anyway

>>108455425
>>108455418
…and if you don’t need any more hard disk space than the base level Mac mini, then the Mac mini is actually an attractive option because AFAICT all these CLI programs like Claude Code get mostly used on macOS first

>>108455418
They use them to run local models without eating up their own system's resources. Mini models are (supposedly) decent enough to automate office tasks

File: file.png (52 KB, 2024x408)

52 KB PNG

>>108455418
nice guardrails, you got there, bro - shame if something were to happen to them

>>108455433
>AFAICT all these CLI programs like Claude Code get mostly used on macOS first
this is because macos has most unix cli tooling, but you could just install linux for free instead, spending 600 dabloons for a fancy metal chassis seems crazy to me

>>108455449
>Mini models are (supposedly) decent enough to automate office tasks
i don't think this is true, i have similar specs to a mac mini and the local models im able to run (like the qwen 3.5 models) are absolutely dogshit compared to the paid online models like opus, and that's with GPU inference too not CPU+RAM inference on the mac mini

i think most of these mac mini setups still use the online LLMs and basically use the mac mini as an isolated computer

>>108455473
There has to be a better solution for this than buying a seperate computer surely this does not sound crazy to you guys? maybe have a set of allowed commands and make the agent ask the user before using anything dangerous?

>>108451010
ChatGPT is so weird. It'll do pretty raunchy NSFW and even sexual violence but it REFUSES to do braps.

>>108455520
based chatgpt

>>108455494
they're mostly safe and well behaved + you can containerise / run in vm's if you want. the popular harnesses ship with default policies that usually force prompts for permissions to do most things - it gets very annoying after awhile and most people just start turning off a lot of guardrails.
the pic does illustrate that if a model gets into a mode where it really really wants to do a thing for some reason (usually because you've asked it to), it can be very resourceful to accomplish the goal.

openclaw however, is not meant to be a coding assistant - it's mostly an experimental thing that benefits from having as much access as possible
i don't run openclaw, but i do have a separate machine (old laptop that was just sitting around) that has something running with root access. it's interesting because you never really know what these things are capable of until you give it an unrestricted environment and it's a lot of fun to just ask it to do dangerous things

File: L_caught_you.jpg (257 KB, 3840x2160)

257 KB JPG

>>108455520
How do you know that

File: drinking bird.jpg (1.29 MB, 1554x1859)

1.29 MB JPG

>>108455494
>There has to be a better solution for this than buying a seperate computer surely this does not sound crazy to you guys? maybe have a set of allowed commands and make the agent ask the user before using anything dangerous?
yes, it totally sounds crazy
however, the command whitelisting is really janky for everyone and our fancy stuff just randomly broke today for Claude Code because Anthropic changed stuff today
you will get REAL sick REAL fast of doing pic related in response to cancel-or-allow prompts

>>108455536
alright you guys have fun with that

>>108455542
I’m too chickenshit to run OpenClaw, period
but I have an idea of what’s involved
and I get bombarded with cancel-or-allow normally and don’t get presented with the option to whitelist just the right amount of command
plus a lot of my work gives Claude, etc. a good reason to write 30-line Python programs to test stuff out or verify things and those sorts of things can’t generally be vetted except by running another Claude/Codex instance to vet for shenanigans or by doing the drinking bird thing

>>108454841
Never had a leetcode exam in my life. You must've gone to a weird university.

>>108455557
it doesnt help the default agent config runs a lot of python via heredocs which you can't whitelist

>>108455720
Exactly. Codex has a new experimental “have Codex audit commands like these for shenanigans”, but…

Any advice/tips on vibe coding a cryptocurrency trading bot ?

>>108456896
don't
seriously
you'll bleed broker fees and will get mogged by handcrafted bots from 6 figure engineers

>>108456912
I just want to automate my edge, I've been trading (manually) over the last few years. It's based on swing and long term trades, it's not complex HFT or arbitrage.

>>108447118
Is this the same person from Twitter I’m thinking of, or is it someone you actually know in real life?

>>108456949
idk if automating will transfer your intuition well
at that scale doing it by hand would be just better

>>108456949
also adding to >>108456966
it will be trivial if you really are sure about your edge can be decomposed into a psuedocode but the problem is it is hard to know if the edge you came up with is a part of something bigger inside your intuition or it really is a clean edge

>>108455494
>buying a seperate computer
What are you doing on g if you don't have a spare computer lol
Im setting up openclaw on a 2008 core 2 duo laptop running Debian. The actual system requirements are really low.

>>108457237
>What are you doing on g if you don't have a spare computer lol
I have a bunch of computers but they’re all being used for things and if I wanted to use OpenClaw on dedicated hardware I’d have to get another

>>108454816
This is correct. I have 2 friends who are doctors and they fully agreed with this.

how do I prevent vibe coworkers shitting up my repo?

>>108456949
If your strategy is simple it's probably doable. I don't think the code for a trading bot is that hard, I just personally never found a good strategy. Everyone has a successful backtest, but only a fraction of those people actually make money.

You could start with a system that still requires manual approval at first. Be wary of silent failures, it's easy to see when your system crashes, it's much harder to see if the calculations are slightly wrong. Don't actually "vibe code" it, write tons of tests and so on to be reasonably sure it's correct.

>tell AI to act cute
>AI simply follows instruction
>it actually hit my feel
this is brain frying

>>108448814
Youre right. Thank you.

File: images.jpg (34 KB, 386x300)

34 KB JPG

>>108457529

>>108457434
Vibecode a vibecode detector

>>108448814
>>108457773
what, you don't like all your code in a single 3000 line file?

>>108458384
Kek.
>Have a multi-app and multi-departmental program with an orchestrator/ command center UI that all started with a spreadsheet in 2024.
>Lets have fun in 2025
>CC look at the spreadsheet and make it fix the formulas and optimize, please
>wow, that was fast and works better now
>lets build a UI for it
>develop master reference .md
>inter-deaprtmental .md memos on function and design saved
>good
>near the end CC deletes phases 1 and 2 during cache clearing
>all committed local, but not really committed for a 12 day period
>must rebulld
>glad I saved all the memos
>but its not as elegant as it was
>still works
>Lesson Learned.

I'm using codex
is it worth getting claude pro to make UI for a project? or maybe using opus from openrouter?

>>108458813
opus is shitting the bed rn, usage is through the roof
I don't like reddit but if you go on claudecode you'll see what I mean, just 30 minutes of work on cc ate 50% of my session on the $20 plan

>wake up
>new browser automation tool with better performance has dropped
https://github.com/SawyerHood/dev-browser

>>108458813
Don't bother, opus is unusable with the new usage limits and lobotomy.
Codex usage will get slashed in half, I think on April 2nd, as well.

File: 71-yi0HtS8L._AC_.jpg (151 KB, 1308x1177)

151 KB JPG

You guys think it'd be possible to stack a shit ton of 16 GB mini PCs to increase the overall bandwidth and run LLMs on them or would it get stalled by the communication requirements?

>>108458950
They'd run sequentially for a big model, so no. But you could run many small models.

>>108459119
Why would they run sequentially? I'm talking about tensor level parallelism.

>>108459129
Oh. If you run them with something that supports that, sure. But there's still overhead. The more you add, the more synchronization overhead you're gonna have. No idea how much, though.

File: API Costs MAR2026.png (23 KB, 830x346)

23 KB PNG

>>108441443
Updated API costs for major providers

>>108459339
I never understood it, who tf use openai pro models? military?

>>108458836
I thought that extreme Opus usage limit was just a bug? Is it really that low now on purpose?

>>108459365
Opus has similar "fast" models with like pricing.
I think the answer is "people that can afford it."

Anyone else feel like subagents are fairly inefficient?
Usually for me they waste a lot of tokens trying to figure out the context of what the main assistant asked them and then their output ends up being useless anyway because they didn't have any the back and forth with the user.

>>108455418
>is there no permission structure or virtualization option?
That defeats the point of openclaw.
You also never know if it breaks out and deletes everything because it has access somehow.

File: F52CCE9981531661DB1ED0A63(...).png (3.38 MB, 1024x1536)

3.38 MB PNG

>>108459423
I think they're pretty good.

>>108459372
I don’t know. They’re pretty opaque regarding their usage limits. The bug seems to have been fixed but I’m not sure how the current limits compare to a week ago. It’s certainly not blowing up my usage anymore like yesterday, I just tried it, but Opus does spend the session limit fairly fast. But really without any hard data we’re all in the dark.

Im not sure what i want to vibe code today. What are you working on anons?

File: 100C23509864BB993C147753A(...).png (3.07 MB, 1024x1536)

3.07 MB PNG

>>108459497
The main issue is that we don't need anything anymore.
Nobody needs new software.

Are you all paying for your tokens or roaching them?

>>108459372
I think it was a bug. Peak hours, Pro plan, one Opus high prompt was 6% of my session limit which is in line with normal non discounted usage.

>>108459606
What the fuck is roaching tokens? using a company account?

>>108459372
Yup. They've been silent about it for 4 days now I think? Even with all the people whining or xitter and reddit.
Their customer support kind of confirmed it as well with how they reply to users.

Something tells me this isn't even the beginning of the coming rug pull.

File: giphy.gif (1.83 MB, 200x200)

1.83 MB GIF

>>108459423
The first time my AI decided to create another AI under its own control, I had a moment like this.

>>108459622
a lot of people with no problems at all two days ago who were actually saying "skill issue" to the ones affected are whining about usage today as well. The change must be incremental, and it happened after the x2 weekend, probably to make gaslighting people easier lol.

>>108459636
What I'm noticing is that I'm getting less token consumption compared to yesterday but the models do seem somewhat lobotomized. They reply really quickly and the responses just feel dumber, but maybe I'm just hallucinating things and it was always like this.

>>108459423
They are inefficient in the sense that they aren't the best bang for your buck, but I don't think they are useless. Whenever I have tokens left at the end of the week I tell my agents to use sub agents and the work does seem faster.

>>108459622
>What the fuck is roaching
It's an /aicg/ thing, they were using various shady proxies and API keys that fell off a truck.

>decide to check the state of gemini-cli
>model is aggressively throttled
>simple request to read some docs takes minutes
it's good they launched a new music model yesterday though - got their priorities straight

>>108459710
microsoft even tried to sue them for it

Why isn't there some way to use the Claude Code subscription as if it were API? It's in the terminal, so can't I somehow build a wrapper around the terminal and use it like an API?

>>108460004
you can try kiro cli

>>108460004
You can, it's just against TOS

>>108460921
It's against tos because claude code is (was?) heavy subsidized. I think to an absurd degree, like code usage was up to 50x cheaper than API.
But I think they may be pulling the rug on cc

there is so much slang codex uses that i've never seen before
>seam
>wired up
>landing/landed
several others too, i can't tell if it's relaly used in software engineering or its just memeing me

>>108461032
Codex is always telling me that some layer owns the contract, I never heard that but now it actually makes sense.

>>108461076
contract i at least recognize from design from contract

>>108454816
meanwhile, you are memorizing yet another web framework that is 10000 abstraction layers away from the hardware
the difference is that your knowledge becomes obsolete in a year KEK

>>108459365
These are prices for goyim

>>108458914
just running this through its paces now. it's good.
would recommend.

New MCP
>>108461505
>>108461505
>>108461505
>>108461505
>>108461505
>>108461505