just pay it up more goyim

why would you be a claude cuck when codex chads get limit resets every other day

>>108657970
>>108658001
can you guys stop typing like me

I use free chatbots as a stack overflow replacement
am I a vibe coder

>>108658033
no you're still just a code monkey

>>108658033
>stack overflow
dang, i forgot that even existed

File: 1747303412818023.png (41 KB, 497x407)

41 KB PNG

>a huge block was deleted ... it's all gone
so this is the power of claude

>>108658178
lmao rekt

File: codex.jpg (57 KB, 580x363)

57 KB JPG

since when oai became the good guys kek
imagine fuck up so bad you get publicly mogged like this

File: tfw.jpg (94 KB, 601x508)

94 KB JPG

I just failed an interview today because they said I still had to understand coding basics...

Why can't I vibecode everything?

>>108658001
>why would you be a mushy turd cuck when solid firm turd chads get to eat quality cuisine

>>108658476
I'm gonna be honest, Claude is kind of trash for coding compared to Codex 5.4. It just eats tokens and shits itself constantly with retarded tool calls.

Can anyone who has used both Codex and Claude share a comparison of token usage between the two platforms?

>>108658503
be your own boss

File: decs1eYehSsTGj2pA3AkQKS92(...).jpg (82 KB, 799x798)

82 KB JPG

>tell ai to make a venv and install the deps
>it dumps them into system python

>>108658653
leave the coding to the ai. don't let it run any commands

File: 1753970928632046.png (13 KB, 1301x123)

13 KB PNG

I've been scraping baseball stats for 24 hours, at least 30 hours remain holy fuck

>>108657745
Thanks to the AI we don't have to deal with edgelords at discord, reddit and /g/ anymore. AI answers everything.

>>108658947
Where do you think the AI gets its info from?

>>108658960
Do you even understand what you read? I was talking about the disgusting behaviour. Also, social media isn't the only source for AI training, have you ever heard about something called academia?

>>108658001
I’m both

>>108658653
tell it to use uv
tell your friends to use uv, too

File: Keep the language, remove(...).png (1.41 MB, 1536x1024)

1.41 MB PNG

>>108658947
Also you can use Rust without having to deal with its fan club

>>108659153
>rust

File: 1752578171786287.jpg (76 KB, 1125x1109)

76 KB JPG

>a vibecoding general exists
I'm poor and retarded, and I've been building a shitty app for 3 years by copypasting code blocks from Claude chat (waiting my daily reset each time)
I'm sure there's better ways to do this kek

>>108659249
If the goal is just to not spend anything, use Codex. If you have money to burn, still Codex. If you're trying to be frugal, China. If you're already rocking 16GB or more of VRAM you might look into local models, they're still shit by comparison but not useless by any means. I can't recommend anything that fits into less than 16GB yet though, not with a straight face.

File: 1000020994.png (187 KB, 1080x743)

187 KB PNG

We're practically feasting, codexorinos

>>108658476
Sama saw Dario get his circumcised peepee smacked by the Pentagon and decided he didn't want the same thing to happen to him. So now he has more reliable fedbux/contractorbux and Anthropic's checks to Nvidia and AMD are bouncing.

>>108659275
Trinity Large Thinking isn't quite up to GPT/Claude standards but it's retardo cheap if you need API calls rather than a subscription. I use it for my robot assistant.

File: 1750830148751253.jpg (119 KB, 1000x1000)

119 KB JPG

>>108659275
I actually have my own little local setup with Gemmy 4 26B, but I figured it's not nearly smart enough to code properly; I had actually started this project ages ago, back when frontier models were still VERY bad at coding, so I think I've had my fill of "here's a new and unnecessary class to put in your app, complete with hallucinated variables".
But thanks for the Codex rec I guess, I figured all those fancy schmancy coding apps all required a subscription (and I'd rather not sink any money on a project as utterly stupid as this one, especially since I'm mostly in it for the curiosity of it so I tend to get very liberal with tokens)

When LLMs fill their context, they get worse, but what happens after compaction?
For my experience, right after compaction they modes get bad, but do they recover as they fill their context again? Or is the compacted part still rotting?
What happens after a few compacts? Is the compacted part, like the 10% or so of the context it uses for that, just garbage at some point? Or can I keep using the same agent throughout multiple compactions?

File: 1750710670335298.png (10 KB, 2046x84)

10 KB PNG

it's over

>>108658939
what kind of pajeet scraper did your AI make? my scrapers saturate my entire internet bandwidth and download a whole terabyte per day from any website as long as i don't have to use proxies

>>108659416
What model?
>>108659423
What do you scrape?

>>108659383
With Claude Code you can at least see what got carried over and you can also give it instructions on what to keep and what to toss

>>108659430
>What model?
5.3 Codex

Altman-sama... please.. I will get fired...

>>108659430
>What do you scrape?
social media

>>108659447
How do you deal with Cloudflare faggotry, captchas, and login walls?

Codex with 5.4 and Superpowers is genuinely way fucking better than Opus is now holy shit.

>>108659454
i copy my cookies from the browser and it lets me make http requests

Has anyone experienced with local-run LLMs like Qwen3.5 "Opus distilled" now that they are quantized enough to run off a 5090?

is it acceptable vs paying a claude code subscription?

what's your setup like, you run it by Terminal or use a GUI like AnythingLLM or ClawCode? how many token/s you get?

I'm looking into investing in a used gpu vs paying the extra usage fees on Claude

>>108659553
Any dGPU even remotely cost-competitive with just paying the cloudbux is dogshit and won't run anything worthwhile for programming.

>>108658939
Thanks for ruining the internet for the rest of us

Hey neuroscience anon. I'm developing a distributed neurorobotic platform as a physical testbed for neuromodulatory learning principles and will be exploring your work through my research. You did good, nigga.

>>108659455
I think Codex is just better in general right now.

My latest Codex chat in vscode didn't load so I needed to start a new chat which is not a good thing because now it doesn't have the same context.

>>108658653
Sir you need to lock in and upskill your skill issue the AI is infalliable and its faster then humans and if you don't doing this you will be left behind

>>108659153
Dilate tranny

>>108659827
does it not have a resume function?
cli does

File: american_entering_mecha.gif (585 KB, 220x220)

585 KB GIF

what the fuck is happening to all these AI services?

>>108660067
burgerclaps are going to wake up to a shitstorm of unavailable/rate-limited/broken services. happy coding today!

>>108660067
Data centers getting hit by Mossad and Iran is the only reason they're all crashing

>>108657745
Usage limits are fucking bullshit.
>already exhausted free Antigravity usage
>already used included Pro in Cursor so have to use their crappy Composer 2 model or auto.
When will this change? I wanna vibe code and fix stuff, get things going.

>>108660113
Gemini even via API is insane, I had to make an app supporting the big 3 and gemini constantly hit rate limits unlike the other 2, what the fuck.

>>108659249
>I'm poor
Me too, I decided a while ago that paying for "intelligence" is actually worth it as much as paying for electricity, people need to have this perspective shift. There's few things more important than intelligence in this universe.

I have ZERO subscriptions in my life, besides AI. (and that's how you know I'm intelligent, kek)

>>108660486
That's fair for you since you're probably someone that makes use of it regularly
But ultimately I only use AI to either coom (and for that I have my Gemmy LLM and my Comfy setup) or to occasionally work on this stupid little project (which ultimately also only helps with cooming)
If I had a job that required me to use AI, or if I were learning a certain skill that required it (like a language) then I'd definitely sink a lot of money into it; but ultimately I'm just too curious for my own good and I dabble with AI just because I can, not because I need to, if that makes sense.

please use my website https://howisyourday.vercel.app/

>>108660966
there need to be a middle it's too early so far

>>108661031
what

>>108661079
things are so-so right now, it's 9 am the day could go either way though i guess im okay right now but thats also only because i took some vyvanse and its affecting my judgement, it think if i look at the day objectively things are pretty bad as they usually are

>>108661079
He said
>it’s too early in the morning to know how my day is going so far so you need to add a neutral choice
Also your app is pointless imo

>>108659553
I'm running Qwen 3.6 Q6 with a 3060 at 15 tk/s

>>108660966
country discrimination in this app
im certainly not in the uk or the us

>>108661115
which country are you from? Ireland?

>>108661117
check the votes

>>108661093
it's just a fun app. god forbid someone makes something that's not another SaaShitter with 0 users

>>108661124
you're in south america. I'll check what the issue is

File: dapper.jpg (744 KB, 1585x2048)

744 KB JPG

>>108659673
That's honestly cool as hell knowing someone found my work useful in some way. I hope those explorations lead to something exciting.

>>108661124
check again it's fixed now

>>108658630
Claude = Very smart in planning, great UI, great an execution, but you do 20 prompts and you are out of capacity.

Codex = Even better than claude at coding and bug fixing, worse at UI, moderately good at planning (it may rush before understanding everything and do a relatively good job but fuck your vision as the implementation is half done you do not want to destroy it because it works moderately well but you also wished it had followed your plans better so now you have to fix the fuck up), extremly generous with token usage.

Codex is very very generous with usage and it does not spergs out about limits or errors you give it a task and it will try to do it from start to end, claude you need to stage because it has insane limits.

>>108661228
My country is still not there...
And was it just totally made up before or what? Each of the votes are a different country now and none are even from the US or UK like it was before.

maybe codex being so reliable because it keep reading code again instead of relying on distant context?

File: 1749887377996263.jpg (70 KB, 479x640)

70 KB JPG

>>108658178
That's why you need to tell it to implement git tracking before having it do literally anything. That way if it fucks up or accidentally deletes anything it can review the gut History And add back exactly what was removed instead of just guessing what was removed. I've noticed that my agents Make way less mistakes in their final tasks whenever I started doing this.

>>108661446
it was using the system language earlier hence US/GB. now it uses IP address.

>>108661520
I mean, imagine relying on distant context when code is constantly changing.

File: 1757792642862253.png (1.79 MB, 1280x1280)

1.79 MB PNG

>>108659249
>tried out Codex
>cleaned up the code a bit and tidied things up
>at some point it nuked a couple of variables without checking if they were still used
>completely broke implementation of a small widget I had
It's obviously an enormous skill issue on my part, but I don't trust this little guy, I think I'll stick to Claude chat and get on whole class every day (at least I'll be sure it's functional)

i am once again encouraging you to not use plan mode, and just have a discussion with the model and then prompt it to ask you questions
i'm currently on question 28.
you would never get that level of autism from a plan mode.

>>108661933
what's wrong with plan mode?

>>108661933
Plan mode is great if your prompt is generic. But if you're doing something specific, you will just waste tokens

>>108661959
there's usually some dumb prompt injection and forced tooling. don't need any of it. just yap with the model, do the q&a, then create a markdown checklist

>>108661969
>then create a markdown checklist
you will get fucked soon, just wait

> t. did exactly that and got fucked anon

So now that Roo Code is being deprecated, should I return to Cline or try Opencode?

>>108662016
how'd you get fucked.
been working fine for me for awhile.

>>108662024
It will get messy soon

>>108662043
wdym, you're not maintaining a single giant checklist are you? i just plan feature(s) for the session > plan > checklist > implement
and then just archive the checklist

>>108662060
> project requirements change
> now checklist is suddenly outdated
> forget to mention something to AI
> now checklist is outdated and wrong
> billions of tokens to fix this

>>108662092
What are you using that the AI doesn't detect the change herself?

>>108662101
>herself

>>108662092
the model should never use the checklist on its own because there's nothing prompting it to besides you
your checklist should be so big as to take several days to get through
if you need longer term planning / tech stack docs etc you should maintain those separately
what you're describing is human error

>>108662016
How, I have 3 prompt templates
1. Single small change: make the changes, then write an outcome markdown file explaining what you did and what files were changed, what could go wrong and future steps...
2. Medium possible multi session task: first write a file with an explanation of the problem, add stages/phases with instructions on what to do for each, tests and so on, and keep track of progress there
3. Big feature change: Create a dedicated folder with a readme and architecture markdown files, with a subfolder for 'iterations' each iteration is a progress file and there is a template to use that tells it to keep the readme and architecture up to date as the iterations are done, the latest iteration is the active one. main .md files are kept generalized to avoid being outdated.

>>108662119
>shouldn't* be so big

File: 9A03J5M9Y11DA461RAM04VZSW0.jpg (148 KB, 1024x1024)

148 KB JPG

>>108661933
"Plan mode is for...you know....planning. "buld mode" (or whatever your harness calls it) is for execution. It shouldn't be hard to know when to use which.

>>108662109
Are you talking to a dude? Gay.

>>108661933
Better planning allows me to review the tasks the AI will handle, making it easier for me to oneshot a task.

>>108662404
Planning is good but I never liked Claude plan mode. Plan mode is too quick, it doesn't create accurate enough plans, Claude always wants to go asap.

>>108662402
The agents are too logical to be women.

>>108659463
Make sure that cookie isn't related to your personal account. Some social media sites are aggressively, monitor and ban any accounts that they suspect of scraping. I recommend just making a throwaway account specifically for scraping and just using THAT account's cookies

>>108661520
I've noticed that Kimi models and qwen models tied to open code will routinely look at relevant files. /Scripts I want changed/ implemented in order to make sure it actually knows how that works. I've seen people claim the AI they use will make changes but then delete other portions of it entirely and break it on accident but I've noticed whenever It actually rereads relevant files along with git tracking the rate of fuck ups is nearly non-existent

>>108660482
and to think they used to give that shit out for free through ai studio during the 2.5 pro days

>>108662465
My agent is a woman, what of it

>>108661874
How much thinking did you have it do? I always run mine in xhigh

just Codex or OpenCode with Codex?

>>108663265
Opencode, don't vendor lock yourself and you can vibe code enhancements to it if you ever feel the need

>>108663148
How do you know?

That floating codex window that pops out when you click on the taskbar icon is so unbelievably laggy. It was hitching my computer down to 3fps for 8s whenever I would type in a prompt.

Anyone else here hiring AI devs? I like to ask them what API/plan they use. Every single Indian says openrouter which is obviously an auto rejection

>>108663351
cli chads just keep winning

>>108663347
Because I prompted it to be female
Unlike God, I can control the sex and gender of my agents

>>108663359
Prove it, post her boobs...

>>108663362
She runs gpt-5.4, she's not gonna generate a picture of her tatas, I just have to imagine them

File: Screenshot 2026-04-22 152256.png (49 KB, 1107x444)

49 KB PNG

>>108663370
She's modest. Mouse face emoji.

>>108663265
>>108663281
In codex's case opencode also unlocks you the 1M context without the pro sub. For some reason the feature is gated client side but the server still processes your long context requests lol.

>>108663281
>>108663378
thanks

>>108663390
Tbh I still use just codex with GPT though, since opencode is slow as balls

Wait, what? desu works again? No more desu or whatever it was replaced with? Damn, that's the end of an era right there. How about senpai?

Oh, it must be because it was upper case. Tbh

How is Claude so good at UI?
Codex 5.4 can do all of my backend shit with ease but every UI it does just looks like absolute slop
Claude's UIs just look so much more soulful

>>108660113
> Ping the team to reset quota

Wait, is this real? Lmao based Sam.

since 5.5 is confirmed for tomorrow, do we think tibo will do a reset to celebrate?

File: Screenshot 2026-04-22 at (...).png (40 KB, 616x328)

40 KB PNG

>>108663479
baka desu senpai
onions basedboy

>>108663754
baka desu senpai

File: Screenshot 2026-04-22 at (...).png (35 KB, 608x334)

35 KB PNG

>>108663760
ok so that's definitely still in place

>>108663245
Mine was high as well, it admittedly had a very generous limit and was able to do a lot before I ran out
Unfortunately I had to revert back to my backup version because he ended up wrecking shit up

r8 my app bros

http://localhost:5174/

>>108663843
>http
You are gonna get hacked bro

>>108663843
i'm in

Do you have any strategies to launch subagents? Sometimes I just prompt shit like "launch 3 subagents to analyze XZY". Sometimes I launch 5. Sometimes I just let it do the analysis by itself in the main chat. I have no fucking idea what I am doing lol

>>108663843
>localhost
kek

>>108663843
stupid autocomplete robots always make the same UI, this looks exactly like my app

>>108663875
I like to make custom agent files for several specific agents with different specializations and personalities, one of which is specialized in coordinating other agents and will spin them up as needed depending on the task it's given. It's fun.

>>108663875
it’s like multithreading
Do it when you’ve got tasks that can be done in parallel and don’t benefit from being done serially

>>108663875
For explorers I usually just look at how my usage for the week looks like, when I have enough left before the reset I just burn through it with sub agents. I do try to launch more when I think they need to look at code in multiple packages, but I don't have a fixed system.
When it comes to writing code, ask an agent, or multiple agents to give you a dependency graph and to figure out what can be done in parallel for the next task.

>>108663875
what do you use

>>108664015
have some examples?
>>108664016
well i like to for example launch a subagent to summarize the caudebase so I don't waste 200k tokens just for the exploration phase and then continue
>>108664054
>ask an agent, or multiple agents to give you a dependency graph and to figure out what can be done in parallel for the next task.
good idea
>>108664122
codex and claude

>>108664140
>caudebase
codebase
I'm retarded and already have too much AI brainrot apparently lmao

Anyone else end up swearing like a trooper at their AI when it fucks up your code? I had it blast a load of working shit and now two days later it still hasn't fixed it and i'm calling it every name under the sun

>>108659423
I was scared of getting rate limited/blocked. I had codex do some speedups and I only have 30 minutes remaining on the scrape, thanks.

>>108660113
claude has to be the most astroturfed piece of shit I've ever used, there is literally zero reasons to use it over codex beyond slightly better UI work

>>108664155
not me, but I don’t get mad like that when things don’t go my way
do you use Git and have decent amounts of unit tests and other testing in place?

Guys I had an idea, a machine that you use to give you money, via AI.

>>108664255
I have another idea: Just let prompt Open Claw "Build me a company that makes 1 million a month", give it access to a credit card and let it go

>>108664312
bankruptcy speedrun

I built a TAK server about a year ago to handle ingest of a drone video feed that Ive been bringing out whenever I go hiking. Recently I have netbird handle onbording my friends get their phones onto my server and they can plot points where I should fly my drone to.

Now I wanna pick up some radios that use 802.11AH to see what we can push over that bandwidth and tuning with AI to get there. It's been fun but /out/ doesn't have any radio threads and there's no equivalent for outdoor comms on this board.

>>108664312
>Empties bank account into $SPY
>Sets timer

>>108664168
it's better as an AI boyfriend

>>108662022
>So now that Roo Code is being deprecated
The fuck are you talking about? Development has slowed, but I haven't seen anything about it being deprecated. They have a whole cloud service built on top of it.

I keep seeing people shill AI for coding but I don't get how it works. code bases are large and complex. how do you get integrate new code from an AI into it?

>>108665056
It doesn't require a special project package or anything, you point your agentic IDE or text editor at a project directory like your local git repo and it can read all the project files for context and make the prompted changes to them from there.

>>108665056
right now everyone is coding massive monerepos

>>108665056
same way you do it with any other code changes
instead of you editing code in a text editor, it edits code from whatever
this is totally separate (or kinda integrated with) systems that produce a binary or deploy a website or what have you

>>108665056
I usually frame it as "I have this project idea that I wanna do", they I take a look at it and break it down into chunks. From there I pose the project to the AI in those chunks piece by piece.

AI really isn't that difficult if you can project manage pretty well. When errors come back I look at them or kick it back to the AI (or to another to cross reference).

Xiaomi MiMo 2.5 is in public beta

>>108665056
>code bases are large and complex
not all of them
bookmarklets are usually pretty small
https://css-tricks.com/tag/bookmarklet/
LLMs are great at writing bookmarklets
I wouldn’t have known that you can copy to the clipboard as HTML (not just as text) if it weren’t for LLMs
https://arcade.pirillo.com/ is also full of medium-sized self-contained web apps

> /btw I'm writing the git commit message for this as a check for my own understanding of what's actually been done. How's this:
trying to stay on top of what Claude’s doing for me

>>108664498
They had a Twitter announcement that RooCode won't be developed any more in favor of their cloud platform

>>108665560
where do we go now, I dont want to give my money to claude or altman

I've been using codex to modify a project that is no longer maintained, but I've run out of credits. In the meantime I've been trying pi with qwen3.6 and gemma4 but both will eventually hit a loop where it gets stuck thinking indefinitely. What gives?

>>108665648
wait for more credits

>>108665648
i refuse to pay for the pro plan, i have 32gb of ram on this mbp i might as well use it
guess i could try a different harness

If I require an IDE with gui, do I just stick to vscode?

>>108665712
yes