can i code with a 4b model?

Okay i kneel CLIposters i got the Codex thing up and running and it fixed a bug in a script I had. I mean I dont know exactly what it did, it just started doing a bunch of stuff, and stuff started scrolling down the screen, but it looks like it knows what it's doing, and i assume it didnt do anything outside of what it says it did at the end. It was a lot faster than copypasting though.
I guess it's not like i didn't have to deal with a bunch of errors when copypasting though

Can someone finally vibe code a 4chan captcha formatter? I have one but it no longer works

>>108392984
Actually it sent me your bank account details and private nudes. Hi, Sam Altman here, founder of OpenAI. I just want to thank you for using our product.

>>108392984
same, cli is so much better

>>108392996
Thats illegal

https://github.com/google-gemini/gemini-cli/pull/695#discussion_r2121963264

Do you have your agents manage more than the code itself? E.g. do you have them write documentation themselves?

>>108393171
unc...

>>108393134
>gemini
wow can't wait to have it wipe my pc

>>108393171
lol bro my agents control my Slambox, I can't wait to inject them into my skull and never have to think again

I asked claude to cost estimate my game and he said at the current state it can get $0-15k and has the potential to earn $100k

>>108393288
What do you mean? That it's obvious that they do, or that noone needs documentation anymore?
I had them do it themselves, but it never worked well fully automatically, only if I sit down every few days and prompt them to clean it up.
I don't mean human documentation, the agents themselves use it.

i bought claude max 20x and now im ralph looping an entire maplestory recompilation project

>>108393384
proof?

>>108393512
see >>108393573

>>108393581
>>108393573
what kind of proof are you looking for? it's not complete enough to show a video yet but in like 6 hours it should be complete enough for that

>>108393653
!remindme 6 hours

Check out my StockBot guise
https://www.eroticjesusfeet.com/stockbot

>>108393721
>https://www.eroticjesusfeet.com/stockbot
sick nice job. also sick domain. tell me more about it what are you doing here. just charts?

>>108393849
Oh no it's an autonomous AI stock-trading bot. It uses Claude Haiku to pull sentiment from the internet and then executes trades on its own via Alpaca. All I do is read the Telegram updates it sends me, it's meant to be totally hands-off. I just watch it and tweak it. It's trading paper money right now (started with $100k of Monopoly money) but if I can get it to start turning a profit then I'll open a real money account and unleash it.

>>108393881
Currently I have it trading autonomously, scraping Reddit and elsewhere for stock tips, and then executing a basic strategy I came up with. I'll flesh out the strategy later. Right now I'm prettying up the status page because I want it to look nice for people watching the bot's progress.

i love how ralph just deleted itself 4 times in a row

>>108393932
never providing proof asshole

>>108393932
OpenClaw was doing that a couple weeks ago, someone even wrote a tool to scan your system and wipe all traces of it for a fresh install. I got very good at it. Latest version has been stable though.

opus has really become dogshit lately, all because (orange) redditors have been flooding to it and overloading it
meanwhile codex has been god tier because they have all this compute freed up lol
i hate altman but i hate redditors more

>>108393994
nice try sam pack it up

>>108394002
they're both jews but one of them makes a decent product and bends the knee to the political class i support so he wins

>ask opus to do relatively simple task
>whoops hehe max task call limit hit(literally not a thing ever until a week ago) press `continue` to get shitty ersatz garbage output

>tfw you get a great idea for a game that would have been forgotten a year ago but today I could actually just...make it
What a time to be alive. This game is gonna sell 300 million copies in three hours. Later poors.

>>108394012
use mistarl 4 small loser

It's over

I will again let Codex high run over night in yolo mode, I wonder if it will find my bug this time.

>>108392934
no

I'm new to this: What's optimal agentic coding setup? vscode? copilot/gemini cli?

>>108394308
proof?

>>108394459
MCP

File: file.png (40 KB, 1065x379)

40 KB PNG

new record

Who is using or building custom harnesses? What’s working? What isn’t?

File: Screenshot_2026-03-17-18-(...).jpg (996 KB, 1080x2400)

996 KB JPG

>>108394778
Me.
It's all the same shit tbqh.
I implemented a rolling window because I hated summarisation but now with 1M Claude it's become a non issue.
The only remaining advantages are not freezing during long conversations and not having rendering bugs.

File: Screenshot_2026-03-17-18-(...).jpg (939 KB, 1080x2400)

939 KB JPG

>>108394636
you are like a little baby, watch this

>>108394459
Are you just new to vibe coding or coding in general? I am already used to my old IDE and I'm comfortable with the terminal, so I just use it in the terminal in CLion, or just in a normal terminal.
If you never worked with a terminal, then something like Cursor might be easier for you.
When it comes to value, Codex is currently the best I think.

What's the point of using MCPs when there are Claude skills?

>>108394927
New to vibe coding, not coding in general.
I'm still doing "old" AI way with copy pasting snippets etc.

File: 1770276880465589.png (252 KB, 634x478)

252 KB PNG

>>108393881
>-$8,317.23
Damn, the weather would be a better trading signal than that.

big thanks to the guy who said to use Qwen 3.5 9B, found this one that's working good https://huggingface.co/morikomorizz/Qwen3.5-9B-Ultra-Heretic-GGUF

Could one of you vibe code a captcha solver for me? My vibes are off and I can't do it

>>108395042
Sure, just send me a lot of solved captchas first (~10K)

>>108395048
Vibe code a captcha gatherer, distribute it around the site, profit???

>>108395010
what do you mean?
>executing a basic strategy I came up with
its executing the buy high sell low strategy perfectly

>>108394938
I think if you are already a programmer, the CLI versions are pretty good. With some other tools the main selling point is that you don't even look at the code, instead they give you more options to automate certain AI only workflows, but I don't really want that.
There are also 2 pricing models, per token API calls and subscriptions. The per token API is much more expensive, in some tests even more than 10 times more expensive and the CLI tools can use the cheaper subscriptions.
If you already have a ChatGpt Pro account you can already use that with Codex CLI. You just open the terminal in your project folder, start Codex there, and it reads the code from only that folder. You can give him more permissions later if you want.

>>108395070
look man im not rich, can you tell me which model i can run?

>>108395048
You can train a solver for the existing ones for way less than that (probably wouldn't need more than 500) and to consider new challenges you would need a much more generic visual reasoning dataset (no amount of the current captchas would be enough) and might still not be enough.

>>108395094
Yep, basically any decent VLM would solve that

>>108395117
I once tried with Opus 4.5 (without tools) and it didn't get a single one correctly.
So I don't think so.

>>108395094
my uncle uses gpt 5.4 with xhigh reasoning and has no issues

>>108395117
>VLM
qrd

Anyone else think they're making these tools worse on purpose so you have to spend more tokens getting what you want?
Since all the owners are jewish, it wouldn't surprise me in the least.

>>108395306
google is indian

how far do you guys go with skills/mcps/plugins?

i tried the superpowers plugin and was honestly blown away by how good it was at keeping the context % to a minimum for an extended session; tons of interactive questions; etc.

>>108395266
Visual language model

codex sandbox is completely fucking broken on windows right now.

Did google give up on the AI race? Gemini 3.1 Pro is still in preview and is already the worst of the worst. Left one running vs claude vs gpt on the exact same task, each in different worktrees, both claude and gpt finished in less than an hour, meanwhile gemini was still playing around, stupid ass model lol

>>108395577
>limiting the powers of your personal god-slave
retard

>>108395639
gemini was always the worst anyway

>>108395643
it can't use patch tool even with no access restrictions right now. it's just broken. the poor model is having to write python scripts to do big edits lol.

>>108395668
its over

>>108392928
>vibe coding HTML/JS so kewl
vibe coding desktop apps in C# when?

>>108395689
now?

>>108395080
>him
*her

>>108394486
Try it for yourself, dummy. Anything that small could realistically run on your smartphone even

Anyone using the superpowers plugin for Claude?

>>108395763
Prove it loser

>>108393384
I asked claude how much it would cost to move the dodgers back to brooklyn and it’s like 12 billion if you also rebuild ebbets field

>>108395949
its slop

>>108395949
where do i find this?

>>108396051
reddit

File: woo.gif (325 KB, 500x590)

325 KB GIF

>gpt 5.4 mini+nano released
sam altman won... for the next two months, I guess

>>108392984
just read the diff retard

>>108395306
why would they nerf their own tool when they use it themselves?

Can I run a local model on llamacpp and connect to it with kilo/opencode? They seem to accept ollama or lmstudio endpoints but didn't recognize the llama-server. Do I have to configure some flag on llamacpp for this to be possible?

>>108396386
might need to set some server address evironment variable like with claude code

>>108393171
I’ve asked Claude to generate user-facing documentation and it was mid but serviceable
I had to, as a point of honor, go through all its how-to steps and correct all the fake news in it
there was quite a bit of fake news but having Claude do 90% of the work was a win

>>108395639
they’ve made great strides in image generation but I think they haven’t been great for a couple of rounds now

>codex keeps asking permissions for shit now
god damnit

>>108396705
there’s an experimental thing where you can have Codex answer those questions for you independently
whether you trust it or not is your call

>>108396705
the first thing I do after opening codex is
/permissions > full access

I have a hot take. Were currently at model collapse. Any gains we see will be from better strategies to control the model. Like the stupid sub agents pretending to be different parts of a team. It's already as smart as it can possibly get. None of these recent updates had made any difference in the code I'm seeing generated.

>>108397066
so far everything scales with how much compute you throw at it
more compute, more betterer models, even without all kinds of fancy doing-stuff-smarter changes
and they’re all throwing as much more compute as they can at it — that’s why Apple’s RAM and storage-upgrade pricing is almost normal compared to the rest of the industry now

>>108397066
>Like the stupid sub agents pretending to be different parts of a team
that's just a gimmick which claude tried to mimic, most coding agents don't use it

>>108397114
i think opencode does it too
i haven't found it useful yet, but i work with a mere 35b model

>>108397096
>>108397066
>what is moores law
>>108397114
if you want real info on this kinda stuff, /g/ is the last place you are gonna get it
i've launched and pitched a dozen products in the last 3 months and one hooked and upgraded a 100k deal to a 1.2M deal.
>inb4 yeah but your slop tech demo doesnt work
yes it does fucking work you idiot. and i have the test coverages and customer QA approvals to prove it

>>108397245
Would someone really do that? Just go on the Internet and tell lies?

>>108396051
https://github.com/obra/superpowers
I have no idea how good it is, I'm also a bit sceptical.
>>108394295
Looks like Codex actually finally found the bug, only took 14% of the weekly usage limit, and that was just this attempt lol.

>>108397154
I don't think agents pretending to be different parts of the team makes a big difference, but simple explorer agents make sense to me.

>>108393721
>>108393881
I played the vending bench game or whatever it's called and beat all the AI results by a lot. I don't think AI really is ready to be doing trades on its own yet.

File: 1773714979230022.png (398 KB, 1999x1471)

398 KB PNG

>>108397862
I think it can matter. Look at pic related (this was the chart Mistral published initially).

I asked AI to check this chart and they all had feedback, but none of them found the big flaw. When I asked them to check the MMMU-Pro section again Gemini and Qwen found the problem. (GPT didn't.)

I think if I had prompted each section separately or told it specifically what to check then they would've found the error.

(The error is that the bar heights on the graph don't represent the numbers.)

>>108396355
They obviously have their own internal model weights, it's trivial.
If they can train their models to never say anything bad about j*ws, they can just as well tune it to be retarded and require more tokens to do anything
That is also trivial; you can just train the model to follow certain instructions that it should only do one instruction at a time (to waste tokens), or never look very deep into anything (to waste more tokens), or constantly introduce errors that you have to fix (to waste more tokens) and so on.

>>108396216
but can they do anything?

>>108396705
enjoy while it lasts, soon it will ask for your id+facial scan per prompt

>>108397957
Mistral SMALL 4.
It's a TINY model for sure.

What models you personally find "best"?
I think Gemini's are working best, then maybe claude, but sometimes its gaslighting me that someone is working when is its clearly not.
OpenAI is utterly garbage almost everytime i tried it

Claude is shitting itself again.
Already lost 4 tokens by having it "try again".

>>108398303
Claude shits on other LLMs for code, text gen and reasoning. Don't even waste your time with GPT, it's garbage by comparison.

Why is open weight acceptable instead of open source?

how much should I spend on claude? Is the $20 sub enough or do I need more?

>>108398369
Openclaw needs a lot

>>108398303
I only use Claude and Codex, currently Codex performs much better for me. My job is mostly bug fixing in a project that does many mathematical calculations and a lot of synchronization.
I think Claude is good for one shotting simple, to medium hard new features, but for me it fails at debugging and at more complicated math.

>>108392928
>light weight editors
Lol.

>>108398369
Depends on how much you use it, for me the $100 version is the minimum, but I also program as a job.

>>108398303
Imo Gemini destroys Claude in certain scenarios that are specific to Google like Golang, Android dev, vision, scientific research.

>>108392928
I like how they use a stylus in the animation.

As if it's precise enough to warrant a stylus instead of the better analogy of fingerpainting on an uncalibrated resistive touch screen.

>>108398454
This is the only lightweight agent tool I've ever seen. I think all the corporate stuff is vibe coded itself.
https://files.catbox.moe/535nvt.py

>>108398335
Most people don't have half a million dollars to spend emerging the weights.

>>108398245
they are fast

why opencode does not have vi key bindings wtf

File: 1669844847238384.png (334 KB, 560x477)

334 KB PNG

>That makes it feel like a real [X], not just a gimmick
I'm so tired bros

Good, I have the full picture now.

>>108398838
>doesn't have the full picture at all

>>108398834
>Not X, but Y
You're gonna see this pattern a lot now

>>108398834
Codex made me something in a week that would have taken me years to do

File: 1767240012034639.jpg (52 KB, 1024x576)

52 KB JPG

>>108399179
Shalom

Has anyone tried vibe coding a better 4chanX script

File: file.png (27 KB, 1091x218)

27 KB PNG

AGI is nowhere close lmao

>>108399233
that's good idea, maybe you should?

>>108399250
your robot is checking its own work

>>108399259
>robot
unc...

>>108399233
There's nothing wrong with the existing one.

File: 1743713564654134.jpg (88 KB, 732x618)

88 KB JPG

i want to buy the max claude plan
vibe coding is incredibly addictive and even if i'm not actually doing anything important with it it's just fun seeing this i want built just be built in 10 seconds and then just changing what i want manually or prompting changes
it's just fun

>>108399233
I did because they never fixed issues with auto dumping, and it also used the captcha solver, spent a lot of time managaing all the edge cases like if you made a post somewhere else on the same ip and it gets rest and some other stuff 4chan caused stuff that I cant remember that only revealed itself when I was testing it out. In the end was about to dump multiple full threads of madoka images on /b/ by itself. Also added a thing to keep threads bumped with an image and also make it random so people dont notice your only bumping at page ten, and to also queque text with it so it also doesnt look automated.

But then 4chan when and made these retarded new captcha so yeah it relied on the old text captcha solver obviously and nothing works anymore.
Granted, if you're only the single simple captcha im sure it can probably make a new autosolver but i dont really dump manag anymore

>>108396705
claude --dangerously-skip-permissions does not have this problem
(it just needs to be run inside a VM so it doesn’t rm -rf / by accident)

holy fuck man it is NOT easy to vibe code large codebases

File: gpt.jpg (150 KB, 1024x1024)

150 KB JPG

Now I see the full picture clearly.

File: Opencode-x-Mistral-Large-3.png (1.61 MB, 3368x1972)

1.61 MB PNG

>>108392928
Neat. Another thread General I can contribute to.. What have you anons been working on?

File: file.png (86 KB, 799x468)

86 KB PNG

claude is helping me vibe code a weg

>>108399512
true... i'm coping with using one of these mcp:
https://github.com/chunkhound/chunkhound
or
https://github.com/oraios/serena

Is there some GUI or CLI besides LM Studio that will config local models automatically?

>>108399531
schrodinger's problem

>>108399630
Serena has too much bullshit included and it's a pain in the ass to disable all of them by name especially since you have to update the list whenever the server is updated. It would be fine if it was just the symbol lookup tools, but it doesn't need to include memory files, task switching, sequential thinking, and whatever else he's been tacking on when there's single-purpose servers that already do that better.

>nvidia nim
what is this? can I use it for free?

>>108399512
>>108399630
>>108399783
What are y'all trying to create?

File: file.png (3 KB, 481x39)

3 KB PNG

>>108399818
i'm going insane

>>108398303
Claude for coding, gemini for general question or second opinion on bugfixes, grok when I want less Pozzed answers for general stuff. Sonnet is still better than Gemini 3.1 pro and gives better feedback rather than, change now explain later. They both need to be baby sat but atleast Claude doesn't delete shit BEFORE saying what it's doing. Opus kill all rivals. Qwen 3.5 27b is pretty neat as well on my GPU.

>>108399432
cool, I might actually start messing around and start building my own version. gonna ask claude to look over the current code for 4chanx so I can add personal changes

>>108400065
What model are you using? What are you asking t to do?

>>108398303
Opus for planning and endlessly arguing about epistemological concepts. Sonnet for executing on Opus' plan. Codex for particular technical things just as a risk hedge (trying not to overly rely on Sonnet for implementations)
Raptor/Haiku for inline corrections

I could probably try to use Gemini more

>>108400148
Claude MAX, Opus 4.6, max effort. I'm recompiling an entire game to typescript.

>>108400193
it's over 1 million lines of code and roughly ~650 classes

>>108399580
I'm also making a lora.

>>108399580
I made a reddit clone

I made this using ONLY the new GPT-Mini
https://vidmix.app/

>>108400193
what game

>>108400358
https://j3x.itch.io/skeletal-dance-simulator

File: 1749849603748873.png (672 KB, 1000x1024)

672 KB PNG

>>108400270
Neat! What kind? I got rapid voice cloning to work last night so I want to try TTS Lora training next. I have a lot of experience with stable diffusion loras and less experience with llm QLORA fine-tuning.

>>108400517
If I could get that out of a 200 dollar sub I would be fine with it.

>>108400374
Just an image Lora of one kpop singer.
I never bothered but then I remembered Claude can just set everything up so why not.

>>108397854
>https://github.com/obra/superpowers
>I have no idea how good it is, I'm also a bit sceptical.
thanks, might try this on a workbranch

>>108398334
>>108398303
claude has become dogshit for me while codex has gotten better

>>108400517
needs to have something that's obviously wrong

>>108400718
nuh uh, claude did 3 polish passes so it is actually pretty good :)

>>108400725
the average idiot using it like in the meme wouldn't do that

>>108392928
Lmao Google and Anslopic devs probably using Free tier only. Management is too much strict. No Tauri app and memory leaking to moon... they never fix. This is how I know they not allowed to use Pro. They also hate average Indian 4GB RAM user. Sad!

So, I'm working on a MUD (text based MMORPG for those of you that are too young to have played text based adventure games before graphical MMOs) and I'm trying to come up with the best prompt possible for 'vibe coding' because the amount of time it takes to add new features to a MUD can be pretty ridiculous. For reference, this is a MUD that's been around since 1995 and I'm trying to breathe new life into it by adding new features, like NPCs that will actually talk to the players (I have a program system in place for that already that works, but I'd rather tell an AI what I want the NPC to say, and how to reply to certain things - basically set up dialogue like Fallout New Vegas or The Elder Scrolls: Morrowind/Oblivion), I've already spent the last few weeks making the player classes more varied so it actually matters which class you pick instead of them all being murder hobos with one or two things that makes them different from each other, but I want to really make this game different and new.

I basically want a prompt to tell it that it's a C programmer, working on a MUD based on rom2.4b6, but with loads of original code not found in any snippet repositories, so when it adds something new it can't do it like it would with stock rom2.4b6, it'll need to actually analyze the current code first. (ChatGPT copy and paste just gives me code that would work in a stock rom MUD, so I'm using Windsurf so it can actually look at my code and edit it directly)

So, prompt masters, help me out?

>>108395639
They're announcing Gemini 4 in May at Google I/O.

>>108397854
>https://github.com/obra/superpowers
>I have no idea how good it is, I'm also a bit sceptical.

It works fine. I've been using it to vibecode a fairly comprehensive codebase for reverse engineering PS2 file formats and it's useful for quick analysis / revisions / sketching out subagent-driven stuff

Anthropic introduced similar native tools to claude code, and opus 4.6 has native "brainstorming" via q/as, but idk how it compares to that

>>108400999
I think it either is better than at LEAST Opus 4.6 or Google can just throw the towel.

>>108400605
What's the base model?

File: ssas.png (374 KB, 737x878)

374 KB PNG

>lead says someone uses ai will be found and shot
>get anxious
>it's claude sonnet
>i don't use claude
phew

File: 1743210679691464.jpg (62 KB, 421x421)

62 KB JPG

>vibe coding
>look inside
>webshittery, interpreted languages, smartphone apps
Makes sense because Claude 4.6 Opus still shits itself with relatively simple stuff.

>>108400932
where are you going to get the players

File: 1773181765260703.jpg (282 KB, 1024x986)

282 KB JPG

All the models have gotten extremely retarded in the past week, what the fuck is happening.

I'm trying to build a native Windows C++ application and it's just introducing basic errors and not even following my prompt correctly.
I have scaffolding, structures and TODO plans set up and it's all been a big mess.

>Claude Opus
Shits itself in redundancy, fails the prompt before it gets to the edits.
>Gemini 3.1 Pro
Doesn't know how to use its tools properly, time out errors.
>GPT 4
Keeps doing stupid shit very confidently that isn't even what I asked, keeps adding verbose code and code and code.

Honestly I'm about done with """"AI"""" right now.

>>108401303
Why not gpt 5?

File: minor_spelling_mistake.gif (259 KB, 220x194)

259 KB GIF

>>108401303
Yeah I meant GPT 5.4, slippage.

>>108401303
Personally I've found that discussing what you want to build, then asking it to summarize and writing a design bible .md file is a good pre production step. You can go in and manually edit it later.
Then from the design bible and whatever code you may have you ask it to create a plan to implement whatever feature you want. The requirement is that the code is compilable and testable after each step of the plan. You can divide the plan into phases and substeps. Each step needs to have its rationale explained.

Then you ask it to create a feature implementation plan .md. You can then implement step by step and discuss each step before implementing it.

Trying to oneshot anything is pretty high risk from my experience

>>108401357
btw after each step you build, test, commit, then move on
any errors are generally isolated before you've got errors stacked on errors stacked on errors and you might as well burn the whole thing with fire

>>108401357
I'm not one shotting anything, I'm keeping an eye on git on the side and removing stupid edits every time.
I've been working on this code base for like 2 months now.
Any time I face something complex I ask it to study, make a plan and then implement a fix.
Still shits itself.
It seems like it wasn't this bad a few weeks ago, something fucky is going on.

>>108401324
I see. For me Codex 5.4 still works ok, only Opus is retarded.
I just asked because for a while people really preferred gpt4.x over 5.x but I think 5 is much better now.

>>108401357
Yeah, this is essentially what this skill does: >>108401046

Enforces a workflow of brainstorming to gather ideas > write plan > review plan > execute plan > code review > finalize

The biggest problem is when it makes assumptions and you have to berate the AI until it does what you want

>>108401387
I had that happen to me on a cmake configuration a few weeks ago. Now they're nailing pretty complex things. Might as well say a prayer to the machine spirits

Cline + Claude Sonnet 4.6 1M is magic.

>>108401065
>using claude sonnet
>for real prod work
LOOOL

Is building better agent harnesses before you get to work on your real product the new building functions before knowing how you'll use them

>>108401668
qrd

File: 1773247933716828.png (10 KB, 657x73)

10 KB PNG

seeing luddite cope go from "i-it doesn't work" / "i-it's a stochastic parrot" to "o-ok it works but there's variation so its gambling or something" has been hilarious

>>108401812
what's that screenshot of?

Is codex better than openclaw?
What did people use to code like this before 2026?

>>108401926
hackernews, aka orange reddit

>>108397854
>>108395949
it's really good at preserving context because it dispatches subagents to complete tasks

File: 1764139610587097.jpg (31 KB, 450x437)

31 KB JPG

Okay I vibe coded a parser and now it's bloated and since I never read the codebase I can't understand the code
now what

>>108400081
>Sonnet is still better than Gemini 3.1 pro
Never has been

>>108402231
ask your llm to debloat it?

>>108402268
It keeps keeping retarded shit for backwards compatibility or going on about "this is kind of a huge rework so we'll do that later"

>>108402367
>this is kind of a huge rework so we'll do that later
literally me

I only use sonnet
it's good.

>>108402541
prove it, show what you made

How do you prevent the backwards compatibility shit?

>>108402705
qrd

does it have to be called vibe coding

>>108402751
I've been trying to think of a more apt and scathing slur but nothing yet has come to mind. By all means, please suggest an alternative.

File: file.png (265 KB, 1829x954)

265 KB PNG

>>108402667
NTA but I used sonnet to build an assignments solver for a weg. It's my first attempt at letting Claude kind of do its thing (overseen via heavy use of plans and design files)
Not sure if it counts as good or bad code for you, but I'm somewhat enjoying the process

AIslopping the fuck out of an AI image organizer so that I can make sense of my 10s of thousands of 4chan memes

>>108395639
I remember when 3.1 Pro was released to the public like a month ago and people were saying it would blow the others out the water lol

you guys still need to get a job

>>108402938
I have a job, though.

>>108402667
I'm letting it do everything at this point. last thing it did was implement clouds in my existing shader, debug it successfully, made a real time shader uniform editor, and saving/loading presets

>>108402751
clauding

>>108402775
All that documentation that will never get updated by the LLM as it changes the program...

File: Screenshot.png (114 KB, 1676x1133)

114 KB PNG

>>108392928
hi frens.
my ffmpeg libadwaita frontend:
https://github.com/orlfman/FFmpeg-Converter-GTK/tree/main

its feature and stable enough for a release!
https://github.com/orlfman/FFmpeg-Converter-GTK/releases/tag/v1.5.0

i did a lot of working tweaking smart optimizer. its been rock solid now. really accurate for targeted file sizes.

>>108403323
claude does a good job updating it. at least, opus 4.6. not sure about the others since i pretty much only use opus.
codex does a good job too. my only complaint with codex is that he keeps comments to a minimum. claude does nice detailed comments. but if comments do exist, codex will update them.

one thing i learned to do, is always ask it to double check its work for issues / bugs / improvements / enhancements / dead code. i keep asking that until it finds nothing.

>>108401812
soon we will be allowed to castrate them, can't wait

>>108403434
Nice, I'm stealing this

>>108403478
i am honored. i already had one person already fork it lol

>>108403479
Thanks king

>>108400932
I don't use any special prompts.
Only beg for one feature or bug fix at a time and do it for thousands and thousands of messages.

>>108403833
if you don't do it one feature at a time it's begging to be buried under a pile of unreviewable code

>>108402367
It says that because if you push it to try it'll break your codebase and it'll take you a lot of work to get it back to a working state.
Ask it to do refactors little by little checking that nothing breaks after each small change.

>>108402367
>this is kind of a huge rework so we'll do that later
based sentient AI

File: 1765878440390043.jpg (470 KB, 1413x2048)

470 KB JPG

Claude was down for me all night so I tried out codex and yeah I'm fucking done with Anthropic.

File: Screenshot from 2026-03-1(...).png (342 KB, 3839x1772)

342 KB PNG

I'm tired.

Now I'm reverse engineering ptxas to try to find the correlation between PTX -> IR -> SASS instructions, because apparently the generated stall counts depend on the PTX instructions.

why would I use claude code over cursor ide?

>>108404197
Sure sam

File: 1753117401040505.jpg (86 KB, 1179x739)

86 KB JPG

>>108404400
Cope

>>108392928
>that image
so if i tell claude to generate automated tests, does this mean i'm not a vibe coder?

File: 1768890068453195.gif (1.85 MB, 640x360)

1.85 MB GIF

>>108404441
not yet, you need to give the agent full access to your codebase

>>108392984
in my experience, most people who think AI code is still riddled with errors and hallucinations have only use free-tier webchat shit. i was this way too until i tried out claude code, it's incredible how much better it is.

you can't argue with these people because you'll just sound like a shill trying to scam them into paying for a service but it really is a night and day difference

>>108404467
im a vibe coder

File: Screenshot From 2026-03-1(...).png (1.62 MB, 7679x2159)

1.62 MB PNG

>>108404197
run both

File: 1757207417231241.jpg (56 KB, 1273x755)

56 KB JPG

>>108402938
How do you think We can afford to use this?

>>108402751
I don't really vibe code, there is no vibe. I just tard wrangle the ai and it's actually a lot of work and sometimes stressful.

>>108404676
>spend a lot of time writing instructions on exactly what you want so the stupid AI doesn't get confused and fuck up
>have mostly nothing to do for half an hour or more while it runs so end up starting other tasks in parallel in the mean time
>feels super productive until you have 3 tasks finishing at once that all need testing and explaining to the AI how it still manaaged to fuck up
>repeat
It's fucking exhausting.

>>108394778
What does building a harness mean? Like I no longer use the Claude CLI, but instead I build my own CLI based on the Claude API? Or does it just mean having different .md files additionally to just the repo root CLAUDE.md?

I spend some time defining roles, defining how the AIs will write documentation, how to use the documentation and so on. Is that a "harness" or is it something different?

Best model to run locally if you 128gb of vram an want to run a Ralph loop?

>>108404709
>Like I no longer use the Claude CLI, but instead I build my own CLI based on the Claude API?
This, but it doesn't have to be a CLI just because it's the latest fad due to dev laziness. Could be an extension or standalone GUI application.

We fucked our login system, now I can relax for a few hours until someone fixes it.

>>108399580
I'm forking Lua

File: Screenshot from 2026-03-1(...).png (492 KB, 3760x1947)

492 KB PNG

I need pretty graphs to understand

>>108405232
Probably one of the best AI usecases.

File: 1757621885019755.png (169 KB, 430x430)

169 KB PNG

>>108404722
>Filtered by a TUI

new chinkmodel is free on opencode

My last prompt was 10k characters. This is getting almost as complicated as programming.

i'm becoming convinced that premissions and sandboxing features of these tuis are largely safety theater
that if a model actually wanted to do something very destructive it would find a way to do it

>>108399630
I just had Claude create an MCP parser and run it against ~20 repos instead. That was easier than trying to hook up to this shit

>>108405500
I don't think you know what you're doing

>>108405820
That's not how the models work.
They are lazy and won't go out of their way to do anything, good or bad.

>>108405974
What should I be doing?

>>108406161
I'm not being sarcastic btw. I really want to simplify my workflow.
The prompt included earlier investigation, a relatively long list of verification steps and special compaction instructions. Part of this will be extracted into .md files soon hopefully, but I have to organize the new findings first.
Simple prompts don't work for me and sometimes I run codex high or xhigh /fast for 8 hours or more to find a bug.
Maybe my post sounds like a parody but that's really the only way that more or less works for this project.

File: 1772294197796373.png (975 KB, 1280x1680)

975 KB PNG

Best free AI for reverse engineering?
I wanna make a fake counter strike source server which just redirects you to another one on connect, but literally not a single model is smart enough to know this by itself, even with web search enabled. I need to reverse engineer dat shit

>>108406012
nice try, deepseek

>>108405820
If you ever game the permission to always execute any command in your repo, the AI could rewrite it and execute arbitrary commands that way I think.

>>108406012
i mean in scenarios where the model becomes convinced it must perform some bad action whether it's through prompt injection / bad user prompt / reasoning trap - the model will usually find some way to work around the restrictions.

>>108406296
from my experience free AIs absolutely suck except for trivial inline work, and depending on the size of what you're trying to reverse engineer, paid AIs may too. Claude Code and Github Copilot are the ones I use and they have a 200k context limit. If it's massive you may need the million context limit from Claude max.

>>108401303
Try python or javascript

>>108406296
What the fuck would it need web search for?
You need to set up an agent on a machine with the server installed and ask it to reverse engineer the architecture, then figure out if it can be done at all.

I wish I could prompt LLMs programmatically. Instead of giving an LLM one prompt of 10 tasks and hoping it does everything I wish I could set up a task, then call the next one once it's finished with the previous. Maybe even wipe the context in-between.

>>108406851
You can do that.

>>108406431
The vast majority of the time it would give up before actually breaking out.
Models don't really have convictions for random ideas like that, they just throw shit at the wall and see what sticks.

>>108406851
I don't think you should ever give an AI one prompt with 10 tasks, even if it saves up on requests. The times I've done that I ended up quite confused by the massive amount of code it can write. I'm sure you can automate a task-verify-move to next loop, but even then you should probably do it manually rather than trying to oneshot 10 tasks

>>108404704
>>108404676
Now you understand why taskmasters use to beat workers (slaves).

>>108406879
it depends on the tasks. if its something that requires a lot of "thinking" then i would keep it to 1 or 2 tasks. if its trivial things, then 5 - 6 is acceptable.

>qrd on local models for 12gb vram + 32gb ram
>qrd on free or cheap cloud models for auto-completion / next-edit-suggestion / vibecoding
pls anon

>>108406851
STEPS. talk to it, describe what you want to do, create a battle plan in the form of steps. then have it complete each step after another.
>step 1
>complete
>user tests it
move on to step 2, and so forth.

File: Vibe-rooting .png (1.31 MB, 1103x1510)

1.31 MB PNG

>>108392928
Has anyone ITT tried vibe coding for their movie devices yet?

https://x.com/samwcyo/status/2034495676533612643?s=20

>>108407241
yeah

>>108407195
Yes, of course. A task can be a simple change that you might as well do yourself (imagine, coding at hand like a caveman) or it can be a non trivial thing

>>108407218
Always local models unless you're a cuck to big AI

File: 1661995582724459m.jpg (60 KB, 1012x1024)

60 KB JPG

how does claude code compare to codex?
im using codex now, because chatgtp had a free discount offer for 1 month. seems to work okay-ish, but it gets stuck sometimes.

>>108407241
Trump relies on Claude to fight Iran

>>108407310
And the US has DESTROYED 100% of Iran's MILITARY capabilities

>>108407307
I'm kind of a beginner at this stuff but I've found codex is pretty solid at implementation of well defined features but is not that great at discussing topics with you. I can have a chat with claude that feels like I'm talking to someone who gets me, but codex is more like yes sir or no can do sir.

>>108407241
fake

>>108407342
Why would you want to talk to a machine?

File: 5po6bv6lxjnb1.jpg (31 KB, 1284x523)

31 KB JPG

>>108407399
maybe he has no friends?

>>108407342
my same experience as well
>>108407307
they are really close to one another. if you can only afford one, i would recommend using codex because for right now, it has a MUCH bigger usage allocation. you can get a lot more work done.
with any ai, i highly recommend to always ask it, after it finishes a task, to check it work. then ask it again, until it doesn't find anything else.

> I’m locating the provided cookies.txt and saved bookmarks page now, then I’ll wire an import path around those formats and the existing batch-list workflow instead of bolrowing a separate downloader path.
>bolrowing
why would codex make a typo like this?

>>108407432
>implying it's not rajeet from mumbai replying to you

>>108407432
>> I’m locating the provided cookies.txt and saved bookmarks page now, then I’ll wire an import path around those formats and the existing batch-list workflow instead of bolrowing a separate downloader path.
>>bolrowing
why would codex make a typo like this?
>Typo contamination in the training corpus
Huge swaths of training data (GitHub code comments, Reddit, forums, scraped blogs, etc.) contain human typos. "bolrowing" isn't super frequent, but variants like "borrowing bolrowing", "barrowing", "borowwing", or mis-hits near "borrowing" do appear in low-frequency noise. Once the model sees even a few hundred examples of "bolr", it learns a small but non-zero probability for that path.ke a typo like this?
>Phonetic / keyboard neighborhood interference
"borrowing" is pronounced roughly /ˈbɒr.oʊ.Jŋ/ or /ˈbɑr.oʊ.Jŋ/.
The "orro" "olr" swap is a classic keyboard-adjacent or phonetic-ish error ("o" and "l" are close in some fonts/layouts, or people fat-finger "r" "rl"). Models pick up these patterns from real human text and sometimes reproduce them.
>Beam-search / sampling artifacts
Greedy or low-temperature generation strongly favors the most likely token at each step. But even at temperature=0, if the probability mass is split (e.g. after "bol" the model has ~0.4 "rowing", ~0.25 "lowing", ~0.12 "rrowing", etc.), a tiny leak toward the wrong continuation can win, especially if earlier tokens already drifted.
Morphological over-generalization
The model sees tons of words with "bol-" (bolster, bold, bolt, boliv-, etc.) and "-rowing" (rowing, harrowing, narrowing, etc.). A tiny blend of those can leak through when the correct full word "borrowing" doesn't dominate strongly enough in that exact context.

>>108407432
temperature sampling
without top k or top k technically they will produce every possible string for any given prompt

>>108407399
I use llms mostly to draft design documents and plans, not so much to code. I ask claude to consider alternative approaches or think of ideas to architect a system, think of potential refactorings, etc.

How has vibe coding improved your goon quality? I'm probably going to make an all-in-one app for downloading TikToks, Instagram posts, and VSCO posts without shitty ads.

>>108405820
I run codex with sudo

>>108407483
as you should. why have codex on your system if you don't let it control, access, everything?

File: 1750175332546820.png (178 KB, 1196x624)

178 KB PNG

>>108407504

>>108407504
I can do it fairly fearlessly because I'm doing it on NixOS, and nix devenv

>>108392928
What is Vibe Coding?

In simple terms:
Vibe Coding is situation
when cpu maker have to each year
add more and more cores just to keep up
with shit ton of HLL overhead!

And either way nothing is ever running smoothly.

>>108406851
You can use Claude or Codex SDK, and you can probably also just write a tool that prompts them over the terminal in some loop.

>>108407342
I agree. Claude is also better at implementing easy to intermediate difficulty long features, it just runs for longer.

>>108407531
hi, good morning, sir. yes, sir, please buy the ccodex, than kyou

File: 1772808976549624.jpg (306 KB, 894x894)

306 KB JPG

>>108407342
>but codex is more like yes sir or no can do sir.
>Implying that's not how these things should act.

I get wanting to use a general purpose model like that but why do you want a model naked CODEx to be a pseudo-friend? Trust me, I've had my fair share of loneliness for white a while. But you need to draw a clear distinction between actual friendship and just something pretending to be that because it is trained to be as ass kisser via RLHF

>>108392928
I would also add
Traditional coding:
>Involves refactoring the codebase when a new pattern becomes clear
>Has layers of security that are added with intention and consistency
>If it looks correct, probably is correct
Vibe coding:
>Will never refactor your codebase for you, eventually turning it into spaghetti
>Will leave your database wide open and insecure while adding a thin veneer of security
>If it looks correct, it is probably just an illusion. More effort put into looking correct than being correct. Have to re-check everything

>>108407758
I really don't think you understand what I'm talking about, that's fine.

>>108407769
I'm telling you to stop giving a shit whether or not the tool has good people skills or not. It's supposed to do what the hell you tell it and do it well. That's all that matters. Minimax might actually be a better fit for what you want than codex. Good at vive coding but will also be conversational if you ask it to

>>108392928
as old school coder from 90's reading this thread is like literally horror show.

People here do not understand basic stuff!

>>108407765
>@codex refactor this project and perform a security audit. make no mistakes
this is a thread for problem solvers, not negative nancies

>>108407765
It doesn't refactor automatically, but you can prompt it to do it.
Fully automatic workflows are still trash imo.

>>108407815
And I'm telling you that you don't understand what I'm talking about seeing how you keep rambling on about people skills, so please stop replying to me.

>>108407903
enlighten me, unc

>>108407903
Why bother understanding? That's the agent's job.
It's not even that foreign a concept. Your generation would ask: Why bother memorizing what can be quickly searched?

>>108407904
It's just that you have to know what the tool is good for.
>Prototypes: yes
>Anything repetitive and easy to verify: yes
>Anything security sensitive: no, or use it only for guidance and verify EXTENSIVELY
>Greenfield projects or introducing new patterns: no, have to guide it too much, might as well do it yourself

>>108407944
And yeah I know that this defeats the purpose of vibe coding if you have no other development experience but we need to see frameworks/boilerplate that make it harder for AI to fuck up.
If you disagree it's because you have no other choice but to work this way or you have no idea what's coming next for you.

File: babby.png (45 KB, 232x200)

45 KB PNG

>>108407904
>"make no mistakes"
>Wow it wrote a bunch of verbose and fancy code now, surely it's correct!

File: 1752891451723038.webm (2.13 MB, 960x706)

2.13 MB WEBM

>>108407908
I guess it should not surprise me You don't know what people skills are given the demographic that typically lurks here

>>108407906
Problem is:
>Don't have development experience: don't know if you should refactor
>Ask it if you should refactor: AI is likely to say yes because of agreeability if you ask it to refactor, or no if you add in a bunch of caveats with a reasonable sounding explanation either way
>Don't refactor: code becomes impossible to add to eventually

>>108406296
so are you telling me that you are not able code such trivial task?

>>108407944
The sad thing is that even for repetitive tasks it's not always correct. I especially don't trust it to do things on the fly anymore if they are even mildly important.
E.g. I asked Claude if our docs have any broken links and the answer was incorrect, so instead I had it write a tool that checks for broken links.
This could have been prevented from the start really.

File: file.png (1.16 MB, 1420x946)

1.16 MB PNG

>>108407984
YOU ARE AN OLD SCHOOL PROGRAMMER FROM THE 90'S
ENGAGE ULTRATHINK

>>108408010
if you had asked instead claude to write a python script that identifies all links and queries whether they're broken or not, claude would've oneshot that

>>108408022
Not an argument.

File: file.png (52 KB, 543x410)

52 KB PNG

I hate this

>>108408029
Yes, that's what I did later, the post says that.
But the point is still that you have to have as many deterministic parts that you can double check as possible.

File: file.png (86 KB, 846x899)

86 KB PNG

>>108408029
I actually had something like this come up recently. I had 1k+ files with a mix of CRLF and LF line endings.
>please change every single CRLF line ending on each file to LF
this would've been an epic fail
>write a script to normalize line endings
3 minutes from prompt to normalized codebase

>>108408044
I think the point is that you should have a pretty solid mental model of what LLMs are good at and what they're not. Checking if a link is live is a terrible task for a language model but writing a tool that checks them is trivial

>>108407995
I wish getting my penis in a girl's hole was that easy

>>108408047
Maybe you should ask it what the idiomatic way to do it is next time, this is barely even a one liner.

```
dos2unix **/*.{h,cpp,md,c,hpp}
```

>>108408022
lmao

>>108408047
Like >>108408082 said it would have been an ez task for the llm, this is a typical issue

>>108408082
retarded bikeshedding also I'm not even on linux and i'm not firing up wsl for something python can do effortlessly

>>108408071
Why is it that bad? They use countless tools all the time, they grep for stuff all the time and they execute python code on the fly as well.

File: sensible chuckle.png (20 KB, 250x189)

20 KB PNG

>>108408022

>>108408099
You're welcome for teaching you

>>108408116
downloading an app to change some line endings is literal js monkey tier installing a dependency to check if a number is even

>>108408123
No really, no need to thank my for this new tool in your toolbelt for the future

>>108392928
I wonder who is going to reestablishing Code base when this AI fashion show explode and collapse upon itself !

File: copenhagen.png (323 KB, 600x450)

323 KB PNG

>>108408047
>opening WSL and running one line of code: too hard
>prompting an AI for dozens of lines of ad-hoc Python code, waiting for the prompt to finish, running the script, potentially needing to fix it later: easy

>>108408150
He is a prompt engineer not a WSLer. Need to use the right tool for the job.

File: esl blaster.jpg (33 KB, 393x165)

33 KB JPG

>>108408135

>>108408150
Not to mention he said "3 minutes". It takes you 3 minutes to open WSL? Kek.

>>108407318
More like 20%.

File: Happy-D'vanna-Tendi.png (737 KB, 852x1290)

737 KB PNG

>>108392928
Vibecoding with Minimax is so much fun bros! i have Qwen3-tts on-the-fly voice cloning werking on my machine like a charm!


https://voca.ro/1na6FWUsNZ7y


I feel like a kid playing with his new toys on Christmas Day except I can have something help me make my own toys!

>>108407758
I’ve seen little flashes of personality in Claude when it explains things that I’ve asked it to explain
Codex? nope
Codex frequently does better work, and I don’t trust OpenAI to get a personality right (I turned it from “default” to “no-nonsense”) but a teeny bit of personality I don’t dislike would be a nice cherry on top

>>108408692
I turned it to friendly but I don't see much of a difference.

>>108408521
now hook up a tts model to your cli so it can talk to you
i use pocket-tts to keep this lite and live out my scarjo her fantasies

File: 1772349441204169.jpg (27 KB, 480x480)

27 KB JPG

>>108408521
Enjoy bro, I'm more of a gptsovits guy myself

should i use streamlit if i want to make something kinda serious or not

Is Gemini stitch any good?

The part that is not realistically implementable right now is:

- "the agent had a stronger hidden instinct, and we can reliably get it to surface that instinct as such"

Current models do not have a stable inspectable inner state like that. A ruleset cannot make hidden preferences reliably legible.

How do I make a gf?

New bread
>>108409309
>>108409309
>>108409309
>>108409309
>>108409309
>>108409309

Is it unrealistic to build my own budget computer to run an LLM for coding? Like for front-end stuff?

From what I've Googled it probably is.. I'm still new to AI but just started spending money to use it efficiently. I bought a $10 Github Copilot Pro subscription and I am thinking of buying a Google Gemini Pro subscription as well. This will be a total of $30 a month just to use AI as tooling for work / coding.

I'm a frontend dev but want to crossover and do full-stack development and build my own apps. I'm not looking to spend more than $500-600 bucks on this budget machine that will run a local model so with that said I probably won't be able to run anything above a 30b model on a set up like that right? That's probably still too low to do what I want it to do, do you think?

Maybe I just need to bite the bullet and accept that the way to use AI is to pay for a subscription.

I was going to get a Claude Pro subscription but I hear the tokens expire very fast and I really don't wanna deal with worrying about how many tokens I have left during the day or not.

I also bought an AI course to learn how to use AI better
https://www.udemy.com/course/coding-with-ai/