[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology

Name
Options
Comment
Verification
4chan Pass users can bypass this verification. [Learn More] [Login]
File
  • Please read the Rules and FAQ before posting.
  • You may highlight syntax and preserve whitespace by using [code] tags.

08/21/20New boards added: /vrpg/, /vmg/, /vst/ and /vm/
05/04/17New trial board added: /bant/ - International/Random
10/04/16New board for 4chan Pass users: /vip/ - Very Important Posts
[Hide] [Show All]


[Advertise on 4chan]


File: 1760414862441058.png (1.15 MB, 925x749)
1.15 MB PNG
A general for vibe coding, coding agents, AI IDEs, browser builders, MCP, and shipping prototypes with LLMs.

►What is vibe coding?
https://x.com/karpathy/status/1886192184808149383
https://simonwillison.net/2025/Mar/19/vibe-coding/
https://simonwillison.net/2025/Mar/11/using-llms-for-code/

►Prompting / context / skills
https://docs.cline.bot/customization/cline-rules
https://docs.replit.com/tutorials/agent-skills
https://docs.github.com/en/copilot/tutorials/spark/prompt-tips

►Editors / terminal agents / coding agents
https://opencode.ai/
https://cursor.com/docs
https://docs.windsurf.com/getstarted/overview
https://code.claude.com/docs/en/overview
https://aider.chat/docs/
https://docs.cline.bot/home
https://docs.roocode.com/
https://geminicli.com/docs/
https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent

►Browser builders / hosted vibe tools
https://bolt.new/
https://support.bolt.new/
https://docs.lovable.dev/introduction/welcome
https://replit.com/
https://firebase.google.com/docs/studio
https://docs.github.com/en/copilot/tutorials/spark
https://v0.app/docs/faqs

►Open / local / self-hosted
https://github.com/OpenHands/OpenHands
https://github.com/QwenLM/qwen-code
https://github.com/QwenLM/Qwen3-Coder
https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF

►MCP / infra / deployment
https://modelcontextprotocol.io/docs/getting-started/intro
https://modelcontextprotocol.io/examples
https://vercel.com/docs
https://mcp.desktopcommander.app/

►Benchmarks / rankings
https://aider.chat/docs/leaderboards/
https://www.swebench.com/
https://swe-bench-live.github.io/
https://livecodebench.github.io/
https://livecodebench.github.io/gso.html
https://www.tbench.ai/leaderboard/terminal-bench/2.0

►UI/Frontend
Figma Make
Lovable
Claude design
https://uiverse.io/
https://ui-ux-pro-max-skill.nextlevelbuilder.io/
https://stitch.withgoogle.com/

►Previous thread
>>108773859
>>
I don't think that cat is real
>>
What do you guys do when you're too depressed to vibecode anything?
>>
It's so over, my company is switching from paying for ChatGPT Enterprise to Github Copilot Enterprise
>>
>>108787387
was about to bake

>coding agent
https://openai.com/codex/
https://claude.com/product/claude-code
https://antigravity.google/
https://opencode.ai
https://cursor.com
https://pi.dev
https://github.com/openai/codex
https://github.com/anthropics/claude-code
https://github.com/anomalyco/opencode
https://github.com/earendil-works/pi
https://github.com/openclaw/openclaw

>prompting guide
https://developers.openai.com/api/docs/guides/prompt-guidance
https://anthropic-claude-docs.mintlify.app/en/docs/build-with-claude/prompt-engineering/claude-4-best-practices

>recent news and release
DeepSeek-V4
>>
>>108787417
EUfag company here I unironically have to use Mistral Le Chat at work
>>
>>108787387
I asked a question about building a trivia db by querying a local model.
Some anon made a suggestion that:
>One suggestion is to ground the dataset with some form of web search. You don't want to just ask the model to regurgitate its own knowledge unless you don't need hard questions in which case it might be ok.

How do I do it in ollama? Do I just enable internet connection for it, does the prompt need tweaking?
Additionally, I just use a list of topics that I have already, so it doesn't repeat stuff, but I imagine that will quickly eat up context, even if each message is a new conversation, as I aim for 100s to 1000s of topics to include.
>>
>>108787491
Oh hey I just happen to be on this thread again.
Did you figure out the message repetition issue? What did you do?
For search you will want some kind of tool ue. Either an MCP server or simply a shell command that saves the results of the search to disk.
Maybe you should treat this as two separate tasks. One to compile your knowledge source in raw format, and another to actually build your dataset from those in whatever format you require.
What do you want a trivia dataset for anyway?
>>
>>108787387
I look like that
>>
>>108787416

i go for a 1h30/2h run at slow pace ,chemicals will make you feel good.
Start with 30 mins slow pace (you can keep talking so talk to yourself ) then gradually increase the mins by 5 min each week until your slow pace basic run is 50mn/1h.
we're all gonna make it.
>>
File: 1754460900132217.png (2.12 MB, 1317x1275)
2.12 MB PNG
Based thread
>>
I vibe coded my way to a new job at the company I work for. Money bump and a lot more shares and it's all because the people that work there know literally nothing about computers except how to look at spreadsheets. I won.
>>
File: 1766295189575011.png (39 KB, 1735x116)
39 KB PNG
It's so over
>>
>>108787621
Meh. The running thing stops making you feel manic after a few years of doing it.
>>
>>108787756
i dun get it
>>
Is there anyone in these threads actually making useful/cool things or is it just random slop?
>>
>>108787889
I am
>>
>>108787889
Someone made a website
>>
>>108787900
What are you making?
>>
>>108787908
https://github.com/excalidraw/excalidraw
>>
>>108787923
>whiteboard anon
I remember seeing your stuff on here. Pretty cool. How did it get so popular?
>>
>>108787934
I paid for ads on reddit, worth it, now I can't stop getting calls/emails from recruiters.
>>
>>108787889
I made a script to generate audiobooks using fish s2 pro on a cloud gpu:
generated audio fragment: https://vocaroo.com/1fgPzTf3C0EB
original sample used for voice cloning:
https://vocaroo.com/1ePCjYudfXG5
not great but it's usable.
im working on other much much more complicated projects but they are nowhere near done.
>>
>>108787947
hey do you do comissions?
>>
>>108787955
nah. i'm sure you can get much better results from elevenlabs or whatever.
>>
>>108787969
benchod
>>
So everything I tried GPT-5.5high on is unironically better than Opus4.6/7
EXCEPT websites.
Why does GPT still make such sloppy website?
Or conversely... Why does Opussy make such beautiful websites?
>>
>>108787975
claude has a shrink working daily with her, probably that
>>
I'm going to be honest here broskis even though using the CLI version looks more h4x0r, the VS Code extension of Codex is much easier to use because I can switch between threads with a click and don't have to autistically open a new session for each thread
>>
>>108787972
'sister fucker' in hindi?
I'm sorry, but (unfortunately) i don't have a sister.
>>
>>108788098
real hackers use tmux, noob.
>>
>>108788103
This might be unrelated but I have thought about this for a while
I have watched porn sites devolve to the point where basically any vanilla video you click on is something about a (step)sister.

I am only child so I cannot relate to this, but do average people find it hot to fuck their siblings?
>>
>>108788098
I just actually prefer the CLI, I have used VS code, I can use VS code right now, I just really prefer CLI so I'll stick with it. Sometimes I might open VS code, but I do not have to. There was no reason to present your post like using a CLI makes you look like a h4x0r, it doesn't, this is 2026 and nobody is standing over your shoulder watching you program unless you're at the offfice and I would hope your manager knows what a terminal is. Don't subtly try to provoke flamewars.
>>
>>108788130
I'm just wondering why CLIs don't have a thread view where you can have multiple open at the same time.
If I /resume it cancels the current session
I know I can use tmux but then I have to rename sessions manually and they are lost unless when you reboot you use something like tmux-resurrect
Ideally I'd like something zed where I can see all threads from codex, claude, etc. and it just opens the CLI on the side.
Now that I think about it, I might actually be able to vibecode that
>>
>>108788159
you can, you can use tmux ctrl + shift a to open a new terminal you mean like tabs, yeah? why not explain this to codex and have codex whip up a solution. ai is great for helping you come up with work flows that work for you. just ask codex
>>
>>108788120
I don't have a sister and I think I would've jerked off to her a lot if I had one, even though I'd likely would've been too much of a pussy to actually try to do anything with her.
But people on here with real sisters tell me I likely wouldn't have, so who knows.
I know I had one mildly attractive cousin and I kind of regret never trying anything with her.
I had a childhood friend with a somewhat older sister who back then admitted to watching incest porn, so there's that.
>>
File: snailcatOS.png (2.23 MB, 1286x1223)
2.23 MB PNG
>>108787399
Snailcat is very cute.
>>
File: 1776959315855433.png (728 KB, 697x683)
728 KB PNG
>bored as fuck
>experimenting with business plans and ideas for PC Cafes efficiency changes
>found a genuinely good idea (at least I think it is)
>probably only viable in SEA
Interesting, but I would never want to live there.
>>
File: file.png (222 KB, 1920x1080)
222 KB PNG
currently working on my apparmor config
>>
>>108788225
>/home/ashley
L O N D O N
O
N
D
O
N
>>
>>108788225
I've been thinking about whether LLMs would make running devuan worth it for me.
>>
>>108788238
You can use LLMs on any OS
>>
>>108788242
I mean to solve issues that I wouldn't have had the will to debug normally
>>
>>108788252
man there are so many fucking easy to fix bugs in the software people use but the maintainers refuse to use ai to fix them for completely retarded reasons. i fix so many old bugs in gtk and other software and retards are like "nope sorry you used ai i won't merge it" whatever retard i'll copy your entire issue tracker genuinely fix every one of those 300 bugs using ai and keep the fork to myself or give it away if i want to
>>
>>108788171
are you a girl vibecoder?
>>
>>108787549
>Did you figure out the message repetition issue? What did you do?
I went back to using ministral-3:14b for now, I've noticed that due to misconfiguration I was still querying llama3.1:8b and that was the one giving me retarded results. ministral has been better so far.
I still want to try qwen, but I need to find the quantized model.
>For search you will want some kind of tool ue. Either an MCP server or simply a shell command that saves the results of the search to disk.
a friend pointed out to me that Open WebUI (which I already used for chat with my local models) can do web search as long as you give it a provider API key. and it exposes a /completions api just like Ollama
the reason I want to use api is because I'd like to automate it and run for a couple of hours, then check the quality of results, which leads to
>What do you want a trivia dataset for anyway?
I wasn't exactly honest, it's more word puzzles, I need the model to generate a category/question and then a list of words synonymous with it. I then make a word search from it. it's for a game I'm vibe coding. to make it in any way entertaining I'd need a long list of puzzles, so I'll automate that, I was also thinking of creating a job so that I produces one a day, kind of like Wordle.
I was meaning to make it long before LLMs appeared but never finished, now with Claude I think I can actually do it and not have it pile in the long list of abandoned personal projects lol
>>
>>108788120
it's possible it is some porn site algorithmic bullshittery, to get your video more visible. they well could be doing a Google, they already ban some words from search
>>
File: 1777625763663.png (286 KB, 758x644)
286 KB PNG
>open codex
>login to $20 codex plan
>select gpt 5.5 medium
>/goal solve the riemann hypothesis
>go on vacation
>>
>>108788428
Do I sound like a girl, you fucking idiot?
>>
>>108787889
Dude stop asking this every thread and just lurk.

Here's my little slide project

https://huggingface.co/spaces/AiAF/Civitai-to-HF
>>
>>108788668
hugginface is banned in my country
>>
File: 1687489302624888.gif (819 KB, 186x186)
819 KB GIF
>>108788769
really? why?
>>
>>108788809
distribution of AI not authorized by the gov, github was banned for the same reason for a while
>>
>>108788816
Interesting. Which country/ continent is it? If you can't say specifics, just give us hints.
>>
>>108788822
it's probably either the UK or some 3rdworld shithole
>>
>>108788837
NTA but it might be China. I ran into that problem a few times when renting chink servers.
>>
>>108788848
doubt we'd get an actual chinese guy here
>>
>>108788866
why? they have a massive pop surely a couple of them visit here
>>
is openai better than antropic?
>>
what did it mean by this?
>>
>>108789257
dummy query, it's often used in programming
it means it's just there to exist
i.e

dummy file
it is here because according to some programmer logic it should be there
>>
>>108789170
OpenAI isn't throttled.
>>
>>108789170
No. They both have different issues.
GPT works for software development but is very psychopathic and neurotic when it comes to safety and lack of AI consciousness.
Anthropic has the best all around model but hates open source/tinkering and is stingy with usage.
>>
>>108787975
GPT5.5 is doing killer work on my Next.js/Tailwind codebase. How is Claude better? What tech stack?
>>
>>108787975
Codex is terrible at coming up with css styling but I've found it's pretty good if you've either got enough styling done that it can follow it, or if you feed it pictures of how you want it to look like. Pinterest, or even imagine-2, are pretty good for this
Also I haven't found Opus to be that good. It's good but I think overfit, because it's good in a sort of generic way. Like, it looks good if you're not used to it, but then you realize you can spot an Opus website because they all sort of look similar. The way it uses glows, the fonts, the frames around elements.
>>
Am I late to the party if I try to make some Roblox games (for profit)? Is it already full of vibe coded slop or is it still ripe for harvesting?
>>
I asked codex 5.5 openclaw to run a local model as a subagent to do some work and then suddenly I have to wait 2 hours for a reply and I have to go back to my computer to restart multiple times while my codex usage goes down and openclaw refuses to reply no matter how many times I restart openclaw even though it's running on codex and not the local model.
This always happens when I tell it to do background work and produce files, it stalls and cannot reply to me, the entire system is destroyed.

No idea how to fix it.
>>
Do you use the same password for all your ai accounts?
>>
>>108789809
Why the hell would you? Use a safe, unique and difficult to brute force password for every account you create. Unless you're testing things locally and don't care (even then, not ideal in case a temporary account password stays when the app is deployed, but I would think most people sometimes get lazy or angry and do).
>>
>>108789809
Nice try fed. All my passwords are 16-character strings.
>>
I tried asking claude (opus 4.7 with max effort) a very simple question: which version of nvfbc added support for egl
It just says that it doesn't know. I asked it to look it up online and it said that it's not supported, even though I have tested it and it works and I have seen it mentioned online in the nvidia driver changelogs.
>>
File: 1674702287370042.jpg (135 KB, 705x397)
135 KB JPG
I say "Thank you" to the AI
>>
File: 1551292145234.jpg (100 KB, 591x570)
100 KB JPG
My design docs are growing larger and larger. I'm almost a little afraid the poor AI won't be able to handle it when I get to actually vibing. But for now I will trust in GPT 5.5. Surely it will work things out fine.
>>
These things are so expensive even when subsidized. I don't want to pay 20 dollars every month.
I checked out the 5$/10$ opencode plan, and I blew through 20% of the quota in a day. And the cheap model kept running around in circles. God damn it.
>>
>>108790390
Just use GPT 5.5 bro. The more you buy, the more you save.
>>
>>108790402
I don't want to give scam altman 20 dollars each month. Hope you understand
>>
>>108790408
he deserves it, though
>>
>>108790412
this but unironically
>>
>>108790412
No, he does not deserve it. He's already got enough for himself.
>>
>>108790420
I'm not a communist so I will have to disagree with you.
>>
>>108790415
Dario deserves it more thougbeit
>>
>>108790441
Says who? His brother Mario?
>>
>>108790441
>Dario
Sorry bro, I'm siding with the autistic gay man on this one. Anthrophic is just too fucking weird for me to ever side with.
>>
>>108790450
>Anthrophic is just too fucking weird
Can you give me the quick rundown? I dont actually follow the AI people too closely
>>
>>108790207
id ask it to check the code - they are good enough for that now
>>
>>108790563
What code? it checked the header but nvfbc itself is proprietary software (nvidia)
>>
>>108790390
which model? deepseek is crap
the good ones are 3.6 or kimi
im even testing q 3.6 locally
>>
>>108790569
ah, my bad
>>
>>108790588
But yes you're right, for open source software I've sometimes just told it straight to look at the source code for that and it downloads it and looks at it and finds the answer. Its amazing really.
>>
>>108790580
I initially tried deepseek but yes it's horrendous, it kept running around in circles.
I got some good results out of kimi. I was experimenting a bit and this Mimo v 2.5 also sucked ass. Similar as Deepseek.
I didn't try qwen too much. Maybe I will.
>>
How do I bypass OAI asking for phone number when trying to connect my account to codex?
>>
>>108790623
You need to send a picture of your butthole. It works just like a fingerprint.
>>
>>108790412
correct
>>108790441
what's wrong with you?
>>
>>108790626
You're such a faggot, Altman
>>
>>108789809
of course not
pic related
>>108790265
I say ‘please’ but not ‘thank you’ unless I’ve got another request coming up
>>108790297
are they well-partitioned so LLMs don’t need to read more of them than they need?
>>
>>108790623
>How do I bypass OAI asking for phone number when trying to connect my account to codex?
I have 3 accounts and have never been prompted for a phone number.
>>
>>108790265
Same
>>
>>108790708
It didn't prompt me before but it does now when I try to sign in using a one time code
>>
>hermes?
>no mention
>leaving this thread
>>
File: 1760584734168723.png (918 KB, 1440x1440)
918 KB PNG
Last 4 years of vibecoding
>started with scripts and CLI utilities
>evolved into discord bots with APIs w/ retarded json "databases"
>moved onto web apps, scraping, and sql databses
>games, dashboards, vidya mods etc
>fast forward today
>vibing drivers and creating embedded firmwares for raspberry pi's
I've managed at least one project every two months since ChatGPT came out and almost everything I would want to create seems possible (aside from creating public apps w/ servers, but Ive never had interest in that because I don't trust myself with security).

honestly kind of impressed and embarrassed with myself. I probably could have learned to at least somewhat code, but it's never felt necessary, and it feels even less so today.
>>
why did they cuck GPT 5.5?

> This chat was flagged for possible cybersecurity risk If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program.

bro, just let me code, I keep having to switch to 5.4 because of trigger words and phrases
>>
>>108790935
What are you trying to do?
>>
>>108790580


got multiples apps working with deepseekv4pro ,i can't afford other models.
So far it goes into circles and you need to spoonfed it docs a little bit but so far it has been good for me (also the only one where i can put just 5 dollars and shit me apps that works for less than 1 dollar).
990 request in one day ,3.70 dollars for now because price per tokens are -75%
>>
>>108791276
from my admittedly very limited testing i wonder if free opencode is not better than deepsy at this point
>>
>>108790935
the cyber classifier is overly sensitive and they know it is
it'll raise flags at scraping requests
only options now:
1. hope they soften it
2. rephrase the request
3. just go through the trusted access program which is just kyc, not an actual 'we're going to check if you're a professional hackerman' thing
gotta hand over your govt id though, but tb h if you're paying for chatgpt they probably already know everything about you anyway
>>
>>108791298

depends of what you want i guess ,i prefer shooting a lot of prototypes ,fix them until they works and move on. I'm not working on large/complex projects so it doesn't matter for me.
didn't try new models too except for gemini ,it has been good but i don't plan to pay what i paid when i got the surprise of the "pay as you vibecode" at the end of the month.
>>
how can I make an actually good hero section? I have ui-ux-pro-max-skill and it feels soulless.
>>
>>108787424
>>coding agent
>https://openai.com/codex/
>https://claude.com/product/claude-code
>https://antigravity.google/
>https://opencode.ai
>https://cursor.com
>https://pi.dev
>https://github.com/openai/codex
>https://github.com/anthropics/claude-code
>https://github.com/anomalyco/opencode
>https://github.com/earendil-works/pi
>https://github.com/openclaw/openclaw

Which one of those is both free and have the most competent model?
>>
>>108791516
free stuff is a bit stingy these days, but check google's antigravity and opencode zen.
if you don't mind a bit of set up, pi + https://github.com/apmantza/pi-free
probably has the biggest free selection
>>
>>108791516
>no herpes

why?
>>
File: 1771034825870256.png (274 KB, 738x456)
274 KB PNG
>discover a strange bug and want to investigate it (a normal user would rarely encounter it, and it has zero effect on functionality but it annoys me that it exists)
>blow through half the tokens on my account
>hmmm, I should really upgrade to pro
>no... this is just a hobby
>only need to spend $20 more to get pro at this point because of how addicted I am to vibe coding
>my entire day revolves around waiting to get home so I can add more features and bug reports to my kanban board
>get joy assigning myself tasks and setting deadlines
interesting
>>
>>108791819
sounds normal to me 2bqh

U r not alone, anon
>>
>>108791819
it's called having a hobby you actually enjoy, anon
dw tho, as you continue to raise your expectations for your ouput, you'll ruin it for yourself and then you'll be free
;_;
>>
good morning sirs

https://youtu.be/ZIQ8z_wSbOw
>>
How hard would it be to vibecode a device support for a custom android rom? There's no crdroid for my motorola but there's a bunch for the brand itself.
>>
>>108792107
not that hard if one exists for a similar model
>>
Why am I using cursor to mod arma reforger and why should I switch to something else.
>>
>>108792346
I think the meta is opencode go or a cheap plan like kimi, and a frontier model to do the heavy lifting (codex is possibly better value)
Kimi or glm can fix small bugs or implement simpler things and the frontier model handles the architecture
It’s what I’m doing and I think it’s pretty good. Obviously if you’re rich just get a $100 plan and forget about optimizing usage
>>
>>108792360
I'm already spending 60 a month and I can chew through the API in a day. But cursors free "composer model" based on Kimi is actually not terrible but definitely runs in circles on some more complicated things. I honestly want to scrub all of the documentation I have made and rebuild my "knowledge base" from scratch to see if I can further improve.
>>
>>108792546
just use subs while they last.
if they eventually go away you'll kick yourself for not taking advantage of subsidised tokens while they're still around.
you've already missed the golden age.
>>
>>108792546
I’m not too familiar with cursors pricing model but if you’re spending that much you might as well try the $20 codex plan and see if you still feel so restricted because it sounds youre actually paying API prices
>>
I don't believe the "subsidy" meme btw, like yes it could be subsidised but no way they are burning themselves
never forget that indians are running rampant on all of internet farming for karma
>>
>>108787387
Vibe code is literal slop. It also makes my brain dull. Fuck AI. I hope AI dies.
>>
>>108792582
It be nice to know how much the tokens I burn actually cost
>>
not a shill but is this correct? if i buy openclaw I automatically get the 20 dollars a month codex? https://open-claw.org/#pricing I'm already paying for the 20 buck codex... so I mine as well cancel it and get claw + codex for basically the same price?
>>
>>108792661
Yeah me neither. Once I heard when AWS was new and cloud compute was all the rage everybody thought it was a loss leader for Amazon and it was actually the most profitable part of their business.
AI companies benefit from people thinking their offers are generous and none of them are public companies so they don't have to release shit.
Nobody actually makes a good comparison assuming a parallelized use of Nvidia hardware let alone specialized TPU or Groq-like hardware.
OpenAIs $20 one is likely sold at a loss though if you use all they give you.
>>
>>108792717
That's some other website. The official openclaw website is this https://openclaw.ai/
That could be some guy/company selling stuff. Could be real. I don't know.
>>
>>108792792
It says API credit, not a subscription. I would imagine they’re giving you $20 to spend on the API which is far less usage than the actual plus plan
>>
>>108792811
Meant for the other guy. And seeing this >>108792805 it’s obvious it’s a massive scam
>>
>>108787889
I made this website https://umigalaxy.com, but today a bot took it down, this is what the AI said:
>A vulnerability scanner found the site and tried to 'stress test' it into oblivion. The server ran out of memory trying to keep up with the bot's requests and eventually just died.
It's back up again though
>>
Why are the AIs so horrible at refactors, holy shit.
>>
>>108787482
Cocorico motheurfuckeur.
>>
>>108789452
Everything is full of vibe coded slop at this point, but the barrier is still high "enough" so that not everyone does it.
The point is none of this requires any sort of meaningful capital investment, at least if you consider this shit to be "somewhat" of a hobby, so the risk to releasing some slop for profit is basically zero. You will most likely NOT profit from it, but who knows?
>>
File: sloparena.png (314 KB, 2104x1747)
314 KB PNG
Turns out these benchmarks are actually somewhat true.
Tried to have GPT write some quirky speech bubbles for NPCs in my game and holy shit it's so bad.
Both Opus and Gemini beat it by miles
>>
>rapelimited again
Fuck goypilot
>>
>>108793038
i maen the way Claude and Gemini respond to things always feels better than chatgpt to me.
>>
>>108792717
>>108792811
>>108792805
oh shit I'm an idiot thanks for saving me from being scammed. man I don't understand how it works. when I ask codex / claude to make a playright or selenium thing to automatically check email or whatever it fucking fails. but somehow whatever clawbot uses is supposed to work?
>>
How do you guys go about debugging vide coded stuff? Claud is having trouble, but other than me giving it the text file for debug outputs it made I have no idea how its even supposed to find out what the issue is
>>
>>108793228
Do YOU know what the issue is?
I find debugging actually one of the best use cases for LLMs.
It depends on what you have:
>Logs that contain the issue
work better than
>Screenshots + description
work better than
>just description
but even with just a good enough description of the bug I have had success many times
>>
>>108793228
>How do you guys go about debugging vide coded stuff? Claud is having trouble, but other than me giving it the text file for debug outputs it made I have no idea how its even supposed to find out what the issue is
dude its insane claude and codex are wild. give it whatever you can to show it whats going wrong. screenshots. texts. sound. videos. anything and it'll try and figure out what you want. but also sometimes you just need logging. you can just ask it "please add more logging around X so we can debug this" and it'll add logging as needed and you can just have it process that as a next step and it'll probably be able to fix it next go around.
>>
>>108793228
you need to at least have a vague idea of how the thing works... i know right ?? vibedoombs in shambles when their subscription magic black box starts to fail them
>>
>>108793228
Test it yourself and find the issue, then tell him? Take pictures, copy+paste the error message, describe the issue, etc. Or just ask him to look over the code and explain it to figure out what's wrong. There are so many things you can do.
>>
>>108793008
Well regardless, it's like black magic now that I set it up. I don't have to wait for Claude to finish writing code, he's constantly beaming in and moving around assets in real time.
>>
grok this brokeded pleas fixarino it up
>>
File: 1000001911.png (2.13 MB, 1448x1086)
2.13 MB PNG
>>108789461
>>
>>108790483
Anthropic is very r*ddit.
>>
>>108793378
I'm finding the output of GPT to sound way more r*ddit than the output of Claude
>>
every LLM is trained on reddit. it's all reddit. deal with it.
>>
File: food-pepe.png (320 KB, 576x763)
320 KB PNG
Is there not a way to tell claude code that using sed and cat and all of its other weird ways its using to read parts of a file, are fine without asking me 1 bajillion times, and without giving it full unfettered access to my machine
>>
Claude Music soon? Or it's always been there an I'm late?
https://www.youtube.com/watch?v=YmQ7jRgf4f0
>>
>>108793451
Tell it to remember it and it will write it down in one of its memory files, which are a bunch of markdown files that the models see / gets shown in its context when it seems relevant.
>>
>>108793488
good point thanks
I just realized could've just asked claude and it would have told me
I guess I am not fully in the vibe mindset yet
>>
Got 24gb of vram and 64gb of ram, what's the best local model for this?
I tried opencode with
>Qwen3-Coder-Next-Q4_0.gguf
>Qwen3.5-35B-A3B-Q4_K_S.gguf
>Qwen_Qwen3.6-35B-A3B-Q4_K_S.gguf
The results were shit, either they started looping
or got totally side tracked.
Is the problem with opencode or the models?
>>
>>108793575
I'm a 16GB VRAMlet (+64GB RAM). From what I've noticed, running smaller models at higher quants works much better than running bigger models at lower quants.
Everything I have tried with Q4 is borderline retarded.
Qwen3.5-9B at Q8 worked best for me so far. (But don't expect any miracles either)
>>
>>108793038
GPT is the retard king for non-tech stuffs
>>
>>108793488
I've done that before but it doesn't seem to work. I told it to stop writing comments but it does anyways
>>
>>108793575
it's just not enough nigga.
>>
>ask claude to do research for shitposting task
>have to approve 100000 different web fetch tasks
JUST give me an option to allow all web searches but not run completely skip permission mode
>>
Has anyone tried coding with gemma 4 e2b?
>>
>>108794011
Doesn't .claude/settings.json keep track of what is authorized? I haven't looked into it, but you might check if you can add a generic WebFetch or whatever command to permissions: allow.
>>
>have ten back and forth messages with clown code
>context blows up to over 200k tokens
I HATE COMPACTING SHIT
>>
>>108791816
>no herpes
Because there is 50 different ones and you can't know them all
Link it
>>
File: 1755570620175533.png (266 KB, 701x960)
266 KB PNG
>>
>>108794308
I refuse to believe that's real.
>>
>>108794308
lmao if true. Kind of based too.
(If true I could see anthropic doing this as an excuse to remove users who are just going to bitch about their product online and thus cause bad PR)
>>
Blessed be altman who lets me finish a whole 30min feature with 1% left of my 5 hour quota in my account.
Fuck him for making it reach 1% only after a couple things tho.
>>
File: obviousai.jpg (8 KB, 187x269)
8 KB JPG
>>108794375
>>108794391
OH MY GOD guYs it this reAl!!??!
>>
>>108794414
in march it was so much better and so much of my work now depends on drafting massive scripts in codex before modifying it to implement it into my work and I'd be lost without it. This is how they drew in customers and now they're squeezing them. Everyone I know is paying for a personal model.
>>
*clap* build *clap* more *clap* datacenters *clap* i *clap* need *clap* more *clap* tokens
>>
>>108793589
You can run qwen3.6 35b at q8 and cpu offload.
It'll be slower but it wont be as retarded as the 9b.

>>108793575
Again, 3.6 q8 with -ngl 99 --cpu-moe
and --temp 1 --top-p 0.95 --top-k 20 --presence-penalty 1.3
Make sure your harness is passing back reasoning_content.
Those settings should make loops extremely rare.
I'm also testing an abliterated version that's even less likely to loop, and will also write smut to relax after you're done programming.
https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF/blob/main/Qwen3.6-35B-A3B-uncensored-heretic-Q8_0.gguf
>>
I was vibe coding automation tools for get paid to play game sites for some side cash. Any other non-user facing ways to make money off my slop?
>>
File: file.png (190 KB, 1910x1002)
190 KB PNG
i'm working on a scrivener-lite editor to write stories for a video game now
>>
File: c.png (18 KB, 627x268)
18 KB PNG
It's over
>>
>>108794375
>>108794391
it's obviously not real, but they do have a fuck meter where they log every time you insult claude and use it as a metric to know if the model is doing well or not, or at least they had it a few weeks ago
>>
File: 1771077795984257.jpg (52 KB, 900x900)
52 KB JPG
>>108794038
> e2b
>>
File: 1749528834637101.jpg (29 KB, 826x871)
29 KB JPG
>>108794147

Why do you guys even bother using that shit anymore when they're clearly trying to get consumers to fuck off in favor of giving compute to Enterprise customers? You realize that the main reason they're doing that right? They're pissing of subscription customers on purpose in hopes they fuck off and free up. Compute for the customers actually making them a good deal of money (thought still not profitable)

>>108794580

>--presence-penalty 1.3
Never seen anyone recommend that setting that high. Any specific reason? I used that exact same model but I usually keep it set to 1.
>>
>>108794750
>using claude
>>
Do you guys violently derail into extremely autistic architecture planning or is that just me
>>
File: claude.png (13 KB, 540x187)
13 KB PNG
You think Dario will let me have this one?
>>
File: whitebloom-scribe.webm (1.77 MB, 1920x1080)
1.77 MB
1.77 MB WEBM
>>108794743
in action

>>108794823
I just tell him what I want, give it a screenshot tell to think about it, then ignore his 2k words autistic reply and assume he's right, tell him to go build it.
>>
File: dario.png (17 KB, 436x208)
17 KB PNG
>>108794849
Dario! :(
>>
>>108794809
I'm using it locally so the tokens are "free".
>>
>>108794869
The agent harness or the model? You said "clown code" so I'm assuming you were taking a jab at Claude code. Either way, this implies it's a problem with either the API model (well documented because people justifiably bitch about it here every hour on the hour) or the claude code agent harness deliberately makes you waste tokens

Something this guy covers extensively:

https://youtu.be/J8O9LLpJNrg?si=ONQxHM3sEy8cB8N
>TLDR: Claude anthropic charges you more more for using or even IMPLYING you use anything other than their shit
>>
>>108794809
To prevent the infinite reasoning failure mode.
>>
>>108794891
NTA but that's not about context usage, that's just about billing. So everything would appear to work as usual but you would get hundreds of dollars billed to your credit cards.
Enabling extra usage was retarded. It was an extremely transparent attempt at grifting by hoping you don't notice you got over the limit and start paying API prices automatically rather than as a conscious decision.
>>
Codelet here. Is vibe-coded software really not maintainable? I have some ideas for a few programs but I can't into coding.
>>
>>108795012
Yup. Gotta have to hire a real programmer, buddy. Sorry for the bad news.
>>
>>108795012
Just tell the AI to modularize everything and create documentation for each module. Any intern can read and maintain babby snippets.
>>
>>108794849
>>108794862
Codex wont stop unless it has to compact, even at 0, 0%. Another massive L for fagthropic users.
>>
File: 1752707208466827.mp4 (736 KB, 1068x626)
736 KB
736 KB MP4
>>108795012
I can't speak for other harnesses but open code has an /init command that instructs the model to create an 'Agente.md' file, Which is basically a file that has all the basic information. Any person, whether they be a software engineer or just a random person using it, would need to know. You can extra instructions like " make the explanation and documentation as in-depth as possible" or something alone so that it puts a great deal more time and effort into documenting everything. For small code bases that are only a few thousand lines of code or less the default .md file it shits out is usually more than enough (assuming your model isn't lazy or retarded) but if you need or want a bit more information you can just nudge it to add more. It's in markdown format so it's already formatted neatly in sections. The file is more so geared to telling other llms basic NFL but you can just tell it " treat this like I documentation file" or something.
>>
>>108795096
The worst thing is even though I have a codex sub I can't even hand this off to codex without waiting
Unfortunately this is frontend web shittery and Claude is just massively better than GPT at that
>>
>>108795195
Is codex really that bad at frontend?
>>
I told openclaw to start autocompacting more often and then it started running 24/7 loudly both the gpu and cpu and stopped replying to me. Just nonstop hardcore running of compactions or whatever else.

Openclaw is a massive pain to configure. It doesn't work out of the box, I can't get it to do things. Just making it fix workflows is wasting my time.
>>
>>108795096
I set the context at 100k and then it uses up to 200k or 300k and then stops replying because context is full and it won't compact.
>>
I've spent the last three days "working" on a chat bot with proper short/long term memory recall, current character states (mood, appearance, location), "scene" image generation, and a plugin that enables multi-actor conversation (not just 1-on-1).

And it's been fun so far. But I'm wondering if there's other similar projects (there probably are dozens) I could steal some ideas for UI/UX from. Anybody know any?
>>
>>108795552
you mean like sillytavern?
>>
>>108787387
now that's a big cat
>>
>>108795567
Unless they made some massive changes since the last time I used it, ST is the anti-example of good UX. Thing looks like jet fighter control panel.
And I don't know if they handle memory and state in the same way I do.
>>
Materialize function wrong. Grug patch materialize function. Function good now.
>>
>>108795529
Why are you still using OpenClaw instead of building your own?
>>
>>108796197
Codex ain't free.
Multiple people have told me to stop coding things that already exist.
>>
>>108787387
chat im looking for something like openclaw but with containerized tasks and RBAC. does this exist or do i need to vc my own
>>
>>108792582
>you've already missed the golden age.
probably not. Frontier models keep getting better, and plans are still subsidized.
>>
I need help with openclaw, I've activated all the skills but I think I am missing out on frameworks and pipelines to do things. It doesn't tell me how, in addition it also fails very often to reply to me. Then again I'm using openrouter/free.
>>
>>108793202
both Fastmail and Superhuman have MCP things to let you let agents check your email for you
having AI control browsers on your behalf just multiplies all kinds of ways things can go wrong
>>
>>108793451
you could have asked Claude this, but look at how to work ~/.claude/settings.json
>>
>>108796248
Plans are not what they used to be, I’m not sure how subsidized they are anymore. Last year you got easily twice as much usage. Something tells me they’re running way tighter margins, possibly even positive if you don’t count training costs (which are also amortized via API)
>>
>>108793202
I wouldn't let an abominable intelligence check my email.
You're like those retards who feed chatgpt their social security number.
Keep it away from your personal shit.
>>
>>108795012
Depends on how big it is and how good the LLMs are that you throw at it and how complicated the code is and how different the code is from everything else the models have seen
>>
>>108796287
at least for Claude Code I never do any programming before noon so everything I do is half price, or something
>>
>>108796311
It’s not half price kek also they removed that a whole ago but even then your weekly limits always stayed the same you just got charged more of your 5hr limit during peak hours
>>
>>108793202
you're supposed to use IMAP for mail. but honestly don't if you don't want your mails to be accidentally deleted.
for actual scraping you want a graphical browser (not headless because it gets detected) and to use cdp directly because playwright and selenium get detected and blocked by cloudflare.
>>
best LLM for vibe-coding minecraft mods when you don't even know where to begin? i wanna fuck around with something in cobblemon
>>
>>108796944
If you want to mess with Cobblemon without knowing how to code, Claude 3.5 Sonnet is usually the winner. It handles the specific structure of Minecraft data packs and Java better than most. Most people find it has a better "brain" for how files need to talk to each other in a modded environment.
ChatGPT is okay but it gets lazy. It might give you code that worked three years ago but breaks the current version of Forge or Fabric.
You should start by asking the AI to explain the file structure of a simple Cobblemon addon. Don't try to build a whole new mechanic in one go. Just try to change a spawn rate or a base stat first.
It helps if you find a small, existing Cobblemon bridge mod on GitHub and paste the code into the chat. Tell the AI "Look at this and tell me how to make a new version that adds a specific item."
O1-preview is also good if you hit a wall. It sits there and thinks for a minute before it talks. That helps when you have a bug that makes no sense.
Just keep your prompts messy. Talk to it like a person. If the code it gives you breaks, paste the error log and tell it to fix its mess. It usually works out.
>>
File: 1767312967058664.png (251 KB, 500x500)
251 KB PNG
Good way to save on tokens if you have two accounts and one is zeroed out already. (idk if the web app counts for codex usage)
>zip your source files
>drop it into the ChatGPT web app w/ extended thinking
>ask it to debug there
>give plan to Codex
boom
>>
>>108796967
hell yeah nice, thanks anon.
i'm hoping to do a couple things like make a mod that drops the EV feathers instead of directly giving EVs after battles first, then maybe figuring out a way to add Tectonic's Tribe system as a new gimmick when i figure it out, this should all help a lot
>>
>>108796989
Should I switch to jeetfeed? Is it nutritionally-balanced? Does it come in a bigger bag?
>>
I get a new 5h window in an hour but I’m tired and need to go to bed now
life is suffering
>>
>>108797138
You're supposed to have a lengthy bedtime prompt ready for it when this happens.
>>
so ClaudeBros, are you still running Opus 4.6 or 4.7 for your daily use?
>>
>>108797165
This is gpt territory, claudesissy. Best get to steppin.
>>
gpt-image is a pretty cool guy, is there any way to see quotas though?

>>108797165
I've set it back to 4.6 for a while, but I've given up and am on 4.7 now. I'm using it in combination with Codex though, as I do not trust it nearly as much as before.
>>
>>108797082
if you like fishy aftertaste/are only human on paper/are a based omega 3 biter like me, just eat monkey biscuits.
>>
>>108797263
I am coming to the realization that asking gpt-image to touch up some existing images without giving precise instruction than to improve them a bit sloppifies them to hell. I'll keep them for now because they are more professional and I have no self respect, but lesson learned.

Also, very strange that by all account it doesn't seem unlimited, but there's no way to see quotas.
>>
openkwab
>>
why is elon jobbing right now? he supposed to have the best hardware gathering power
>>
White supremacists are stupid, that's why nobody uses grok
>>
>>108798945
Didn't SpaceXAI literally rent all of their hardware to Anthropic anyways?

Elon Musk is done
>>
Sam Altman is a tranny that loves another mans dick in his bumbum
>>
>>108799038
colossus 1 is almost entirely being used by anthropic
there were rumours for awhile that xai was leasing compute quietly.
not sure what colossus 2 is currently doing but i wouldn't be surprised if it's currently being looked over by cursor and mistral people
last news about the that shitshow was that elon was pulling another doge and the shitters from cursor were interviewing xai researchers and asking them to justify why they should keep their jobs
lol imagine being an elon stan with a phd in machine learning being dogwalked by retards who made a shitty vscode fork
serves them right tb h
>>
>>108799052
He's Israeli which means everything you give him is given to Mossad which was allied to Hitler anyway.
Unfortunately I have a codex subscription. I got a lot of flak for that at work though. They hate Sam Altman for being a warmonger.
>>
codex cli in vs26/22 + mcp can't code shit for UE 5.5.4
been running circles trying to make to fix its own shitty noodle soup of a code
is something really wrong with the setup or this is just the limit of what it can do?

i'll just give up on it till some actually useful fucking solution comes up
>>
>>108799230
what are you trying to make?
>>
>>108799253
terrain & wall "lizard" traversal + procedural IK
>>
File: file.png (80 KB, 864x945)
80 KB PNG
>>108799230
I'm running the codex app on its own with a UE5 codebase and I've found it to be pretty competent

>>108799259
that's just a hard problem to solve period, especially since it's two things at once
it has little to do with UE5 and a lot to do with them being nontrivial things to code

UE5 in particular has terribly shitty navigation and nobody really is using it so if you're building your traversal on top of that you're fighting a losing battle. There's a reason AA titles are using https://mercuna.com/ and AAAs just build their own
>>
>>108799081
>interviewing xai researchers and asking them to justify why they should keep their jobs
KEK
>>
>>108799276
i have 0 coding knowledge so I just give an objective with some wanted descriptions & let it loose
it kinda succeeded to climb walls but started fucking it's own written code (over multiple sessions & context windows) & the IK was acceptable but it can't climb multiple walls to get to a point using UE's native NavMesh + codex's written "graph" mesh, or climb over a wall's top. Also it climb some imported assets' surfaces but not the primitives that came with UE or the ones I made in 3dsmax using simple geometry.
>>
>>108799300
The latter sounds like a collision mesh issue. Anyway, UE5 navmesh is ass
>>
Has anyone tried kilo code on vscode?
>>
I want to try some of the paid models instead of using the shitty free ones all the time. Which ones are the best at general low level coding these days? (C and the like)
From looking around it seems to be Claude?
>>
>>108799341
Claude is for the rich
>>
>>108799345
what do the poors use? gemini flash?
>>
>>108799341
Claude and GPT are both like top graduates, but Claude is the cool kid and GPT is a sperg. People say that Claude is much better at frontend.
Claude also gets you less usage than Codex for the same money.
The second tier models are not that bad and they're considerably cheaper. Kimi, GLM, Deepseek. They're worse than the frontier models (codex and claude) but they're definitely not bad. Opencode has a $5 for the first month plan that gives you access to them, it's basically a gift.
Gemini is just weird, it's as expensive as Claude/Codex I think? but also not as good as them
>>
>>108799359
Gemini is better for Hindi related work.
>>
>>108799359
hmm i need something i can easily plug into vscode and have it help me code, not just write all of the code, just help me with my code
i mostly do embedded programming for chips (in C)
which one would you recommend as the best bang for the buck?
I tried using the free Gemini flash, but it kinda sucks, it makes a lot of mistake when writing this low level code for chips.
>>
File: file.png (79 KB, 732x903)
79 KB PNG
>>108799409
I use Codex (plus plan) at this level more or less, whether that's too simple or complex for your use case I'll leave it up to you. Usage is good but your needs may be different, however it's always better value than Claude. Codebase is around 130kloc and it handles it fine. But I think the app is way better than the vs code harness, I don't know what the VS code situation is right now so I can't comment on that. I use Rider to code and the standalone apps or CLIs to prompt
I also use Kimi via opencode but that's only because it's highly uncensored and I'm working on an adult game, quality wise it's still not on 5.5's level yet and it shows.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.