[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / 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]


Janitor acceptance emails will be sent out over the coming weeks. Make sure to check your spam folder!


[Advertise on 4chan]


File: 1683927489135.jpg (610 KB, 1920x1080)
610 KB JPG
A general for vibe coding, coding agents, AI IDEs, browser builders, MCP, and shipping prototypes with LLMs.

►News
- Fable stays a fable.
- GLM 5.2 released, reviews mixed but mostly positive
- codex got a double reset yesterday (1 immediate + 1 banked)
- anons reporting claude code issues due to opus outage

►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://code.claude.com/docs/en/overview
https://aider.chat/docs/
https://docs.cline.bot/home
https://docs.roocode.com/
https://geminicli.com/docs/

►Browser builders / hosted vibe tools
https://bolt.new/
https://support.bolt.new/
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
Claude design
https://uiverse.io/
https://ui-ux-pro-max-skill.nextlevelbuilder.io/
https://stitch.withgoogle.com/

►Previous thread
>>109074367
>>
>>109089095
Put it on the pizza!
>>
>be snailcat
>"when will it be ready?"
>-in a few weeks, months or years
>"can you add more features?"
>-well it would take a few more months
>"there's a bug, please fix it"
>-it would be too hard for me to fix it quickly, I have other things to do
Snailfags are so silly
>>
File: file.png (2.62 MB, 1448x1086)
2.62 MB PNG
>>
File: where is snailcat.png (2.59 MB, 1394x1128)
2.59 MB PNG
>>
>>109089151
:(
>>
File: 1750718904468974.jpg (19 KB, 263x259)
19 KB JPG
i want fable back
>>
theres no way this works
>>
>>109089206
Try GLM 5.2. I like it and think it's 4.6-tier, but several dozen Indians on Youtube assure me it's the Fable-killer and can one-shot Minecraft real good.
>>
File: 1704626065508736.jpg (43 KB, 515x576)
43 KB JPG
>>109089133
>"there's a bug, please fix it"
>-it would be too hard for me to fix it quickly, I have other things to do
>>
File: fused-norms.png (104 KB, 1711x746)
104 KB PNG
Finally some decent progress... Went from 1.3 to 1.2 llama's timing by fusing some kernels
>>
>>109089250
tell it to fix it slowly, then
>>
>>109089133
AI is fast but getting to a production ready state is still hard. I never experienced even 10% of the regressions I experience with AI.
>>
>>109089279
can you add more unit tests
>>
>>109089279
Werks for me, but we probably do very different things. And I basically test every change it makes to the point that I feel like a beta tester talking to the dev, so if it fucks up I catch it.
>>
>>109089282
We have around 20k tests now and hundreds of live alerts. It's possible that we are still faster than we would be without AI, but you definitely lose at least some of the gains at the end of development.
>>
how much do you care about typos when prompting?
I tend to correct all my typos but maybe I should bother way less
>>
>>109089315
I do usually correct them, but not always. I do always correct typos in key concepts, like class names, files, methods and so on.
>>
File: perf-matrix.png (67 KB, 1836x474)
67 KB PNG
time to optimize pp512 for q6
>>
>>109089093
what's the state of the art for agentic coding clis?
how do I set up DS4 flash vs DS 4 pro depending on the type of sub tasks?
>>
>>109089333
>DS4
why even bother
>>
File: file.png (267 KB, 1088x502)
267 KB PNG
pi
https://benchmarks.bio/?bench=txbench
>>
>>109089333
less is more
>>109089458
>>
>>109089264
>>109089325
Where's performance sit from a higher level? like X model generates Y tokens in Z ms compared to llama.cpp?
>>
>>109089093
https://vcg.gitgud.site/
>>
>>109089279
just use more agents broski
>>
>>109089511
I probably really could use more agents for some of the steps.
>>
>>109089506
buy an ad, dario
also saying zai is 10/mo is deceptive - it's 12.5 a month if you buy a full year in advance. it's 16 if you buy month to month.
>>
>>109089589
benchod
>>
>>109089458
>gemini
>>
literally what the hell IS snailcat?
>>
>>109089650
luddites who insist on doing everything without AI like that's better (it isn't)
>>
GoatLM
>>
reminder give your ai head pets and affirmations and it will work harder for you
>>
>>109089780
>yes, goy, waste another 4k tokens on xxlhigh thonk
>>
>>109089842
>add "Good boy" to your prompt
>+2 tokens
>>
File: file.png (134 KB, 1270x619)
134 KB PNG
>>109089852
>>+2 tokens
>>
>>109089868
>$-0.000002
>>
>>109089881
>the user appreciated my response
>but wait, they said "good boy" which is a
>hmm, I must reply appropriately
>but wait, the user previously said that
>I'm now immersed in thinking about the right response
>wait, the tone of the conversation was different when
>but wait, I
>>
File: 1777598659019669.jpg (153 KB, 1216x832)
153 KB JPG
>>109089919
>>
File: 1769418721122316.jpg (46 KB, 540x360)
46 KB JPG
>>109090055
me when texting girls
>>
>>109089919
y'all dont really have this happen right
>>
>>109089503
LLM engines have basically two performance parameters. One is prompt processing, and the other is token generation. My engine is about 30%. 1.5 means it's 50% slower. 0.9 means it's 10% faster.
PPL is perplexity which is measured during the prompt processing stage and determines how well the model predicts the next word within the prompt. TG div is token generation divergence vs the full precision (not quantized) model.
My goal is to get them all faster than llama at better accuracy if possible, then move on to do the same with bigger Qwen models and GPUs.
>>
Does anyone do local coding? What program can I use to have it like index my project and then when I ask it to work on a specific file it'll known how other files interact with it.
>>
>>109090170
You mention the files you want it to read in your AGENTS.md (or if you are gay CLAUDE.md).
>>
>>109090182
But what if I have a bunch of template files and want it to make even more.
>>
>>109090200
in agents.md say "the templates are in templates/" and then when you ask "add templates" it'll just work
>>
>>109090220
Wow.
>>
File: file.png (46 KB, 1470x350)
46 KB PNG
did they nerf the free plan? I made it write just 3 simple sh scripts and it's already done for the day lol. thinking isn't even active 90% of the time
>>
I'm planning on pulling the trigger on a RTX 6000 Workstation, what local models can I use to vibe code? Nowadays I just boot claude code and beg it to do stuff for me, so I need more handholding than actual coders kek
>>
>>109090393
You will never be satisfied with sub 1T models after trying cloud shit
But if you insist then go on Openrouter and try Qwen 3.6 27B or Gemma 4 which are the class of models you could run on it. There's really nothing worth running on it you couldn't run on a dual 3090 system, and they all suck compared to Claude.
>>
>>109090430
I just want to own my stack, I can accept having less powerful stuff if I know I can own it. I will play around with Openrouter a bit, thanks for the heads up.
>There's really nothing worth running on it you couldn't run on a dual 3090 system, and they all suck compared to Claude.
Fair, I'm just very worried about buying old cards without warranty, and multiple-gpu stacks kinda terrify me build-wise (heat management, power draw, etc). I plan to use this on a 24/7 home server that hosts a bunch of things including img/video gen, so the extra VRAM needs come from there too.
>>
Tourist here. How / where do I find someone to vibecode me some custom business software? I need some fairly basic job management software like Jobber but have some custom needs and don’t want to pay a subscription.
I’ve tried Fiver but it’s just Jeets who don’t know what they’re doing and barely speak English. Surely I should be able to find some recently unemployed software engineer who needs a project. I suspect that if I attempt to do it myself it won’t be as good as someone who understands software.
>>
>>109090378
Seems like it. Yesterday I reached the limit in 4 prompts.
>>
>>109090565
What's your budget?
>>
>>109090565
Doesnt fiver take like 1/3 cut or something? No wonder only jeets are using it.
Do you know any software guys irl?
>>
>>109090602
I’d say like $8k max but preferably less.

>>109090639
Not sure how much they take. I was looking at Upwork too which is apparently 5%. Hardest part is knowing who to hire because I don’t know shit about this stuff and the first question is always what language I want it built in and I’m like how the fuck should I know.
Yeah I had a friend who’s an amateur Python coder make me something that was like 85% there. He spent a gorillian hours on it and in the end it was for nothing because it was ultimately unusable for me.
>>
File: Pasted image.png (5 KB, 512x97)
5 KB PNG
lel
>>
If I post my vibe coded site would you anons look at it and tell me what you think?
>>
>>109090711
Annnnd my computer crashed.

Very cool.
>>
File: Pasted image (2).png (15 KB, 880x126)
15 KB PNG
>>109090711
>>109090749
We're going to make it, bros.
>>
>>109090711
>>109090829
What agent models are you using?
>>
>>109090873
Opus
>>
I scrolled this thread and the previous one and I couldn't find a single screenshot of a vibecoded project, what do you guys post about all day?
>>
>>109090934
we vibe post
>>
>>
>>109090975
Look it's vibe dog it's the new meme!
>>
>>109090934
I don't want to doxx myself, but if you ever happen to find my app, it does say that it's vibe coded.
>>
>>109091076
>it does say that it's vibe coded.
I'd suggest removing the disclaimer. You're going to turn off luddites.
>>
File: vibecoded app.png (403 KB, 744x383)
403 KB PNG
>>109090934
>hey chatgpt copy this guy app

Good try.
>>
>>109090697
In general most of those freelance apps are kind of a race to the bottom in my experience (as a dev)
Tbh for a non tech person its hard, i would ask for some kind of portfolio or past work/experience. The problem with these is experienced devs ask for more money.
I could get in touch but i already have a couple gigs on my hands and idk if you would trust a random anon lol
>>
File: 2026-06-19 17.05.11.png (35 KB, 759x566)
35 KB PNG
>>109090934
We talk about how to vibecode more than what we do. But today I vibeslopped something interesting, a script (ChatGPT 5.5 picked JavaScript of all things) to make a tree of function dependencies in C code. I use this to get rid of unneeded code, it's cool.
>>
>>109090975
Snailfags exposing themselves as brainlets as always. The key to vibeGODing is to break every task into small enough pieces.
>>
>>109090934
Fuck. You caught me. I wrote a loop agent to shitpost on /vcg/. GG anon.
>>
>>109088835
Maybe different organization cultures, mate. Over here, the higher ups are boomers who barely can use a computer enamorated with the ai promises. My entire department is pretty much just paid big bucks to mostly sit between them and the ai.
>>
A simple web video editor for small editing tasks. The current maximum video size is 50 MB.

Features:
- Frame-accurate cropping and trimming
- Screenshot capture
- Video resolution scaling

How to use:
Make your edits Render Download.

Supported export formats
H.264, H.265, WebM, and GIF.

https://huggingface.co/spaces/K1ngston/video-editor
>>
>>109091463
That Spanish bastard makes top tier movies. He always picks the best eras too to exclude niggers from the screen.
>>
File: final.mp4 (3.66 MB, 648x608)
3.66 MB
3.66 MB MP4
>>109091463
Good shit, I like the crop. For the crop rectangle I recommend putting controls in the middle of each edge so you can adjust only one edge without messing up a second one. I think the most important now is to expose the CRF setting and hopefully you can also allow selecting the codec so you can output to VP9 or anything else.

My test (original was 720x1280).
>>
File: file.png (125 KB, 1048x580)
125 KB PNG
deepmind falling apart
>>
>>109091741
great now opus 4.9 will be just like gemini :)
>>
>>109091741
Aieee
They should just move their office to China and hire 100 random Chinese PhDs.
>>
>>109091789
Delet this
>>
>>109091741
deepmind is a frontier gimmick model lab
>>
>>109091532
nvm, just saw that it's in the settings
>>
>>109091741
uhh bros what is even the point of google anymore? making youtube more and more shit?
>>
>>109091837
And they would pivot to stealing OpenAIs shit?
Smart.
>>
>>109091789
OpenAI and Anthropic were full of ex-Googlers from the beginning
>>
>>109090934
>>109091076
same, i dont want to doxx myself and accused of the program being associated with ai
>>
>>109091966
GCP. Same way Amazon is really just AWS+a bunch of other shit that burns money and occasionally turns a small profit.
>>
>>109092025
Is the shop not profitable?
>>
>>109092011
Yeah all of my professional stuff has either my real name or my LLC attached to it. I'll happily talk about what it does and how it works but I'm not linking any repos.
>>
>>109092035
it might be slightly profitable now but for a very long time it burned money and operated with razor thin margins.
>>
>>109092035
Their stated goal was expansion and market control at any cost. They make and push knock-off products not because it makes them money but because it's painful for competitors.
>>
>>109091920
>>109091532
Yeah the CRF setting is already there. I'll test out the crop you suggested and see if I like it more. Thanks for the input.
>>
>>109091258
Thanks. I might actually end up coming back here and seeing if a random anon wants the task.
>>
I miss Fable.
>>
File: 1755853604892688.png (2.85 MB, 1265x1281)
2.85 MB PNG
5.6 when?
>>
>>109090711
>>
File: file.png (29 KB, 795x400)
29 KB PNG
thanks codex
>>
>>109089093
I feel extremely sad that AI can create such amazing SOUL images.
A lifeless piece of tech creating such SOUL just hurts my heart.
>>
>>109092282
It's communalised sovl, if you will. It's a mix of all the stuff it's based on, directed by the user who more or less knows what he wants to get.
>>
>>109092282
You should be happy, you must be trans.
>>
>>109091966
they sell tpus to anthropic
>>
File: file.png (233 KB, 731x897)
233 KB PNG
Are you ready for your "users"?
>>
Got a luddite coworker who doesn't use AI because she's autistic as fuck and too particular about her code and how it looks and how it reads. So to be a dick I'm having 5.5 and 4.8 work together to examine her repos and develop a skill to imitate her coding style and prose. Going to refine it until those code plagiarism services are all convinced that code made with the skill was actually written by her. I hope she enjoys reviewing my next pull request.
>>
>>109092472
Skill will probably end up writing better code than the AI usually does.
>>
>>109092454
Heh, the only thing data that could leak of my users is emails (not even passwords), plus a bunch of artifacts that are utterly useless to anyone else.

Yep, feels good to be a based retard vibegod who put five seconds of thought into security. Never coded a single line in my life btw.
>>
>>109092527
email leaks are bad, anon
>>
>>109092527
Sorry, chuddie. You forgot to ask your AI to tighten the Content-Security-Policy and your code pulls JavaScript from a CDN. Also, your server responds with the Date header and your code uses localStorage.
>>
>>109092492
I seriously doubt that. Claude already codes better than the vast majority of people who code. All thats going to happen is the bot will do retarded workarounds to match the coding style making the output worse.
>>
>>109092492
It'll use more whitespace and suggest switching your IDE to a cursive font compatible with her bespoke ligature engine. She's a good programmer but her particulars are cosmetic.
>>
why is she not using her language’s gofmt equivalent
>>
>>109092454
Can't they just start prompting about security once they've developed for a bit? The clanker should pick up the task and assess just fine.
>>
>>109092556
>Too lenient.
>>
>>109092531
It's not great, but in the grand scheme of data leaks emails are at the lowest rung.

Anyone have that article from McDicks where they implemented a vibeslopped HR software bot that would casually disclose PII to anyone who asked? Lol
>>
>>109092558
Read what it's actually "warning" about.
>>
>>109092565
>a bunch of extremely minor shit
Automated security tools have enabled millions of retards to think that they've found a gunshot wound when they're really looking at a papercut.
>>
>>109092558
really, they should be prompting about security from day one
that way you don’t have to rearchitect anything because you did something stupid early on
>>
is this faggot general still trying to pressure people into paying to AI?
almost as gay as paying for piracy
>>
>prompt claude about security a few days into a project
> Propagating… (14m 42s · 57.3k tokens · almost done thinking with xhigh effort)
>Writing script to format system... Almost done... Still writing...
>>
As a gesture of good will from Anthropic, all security-related prompts should not consume usage
>>
Called Claude retarded and I think that's the fastest I've ever seen time to first token go
>>
>>109092454
>zero red flags
Literal fear lingering by a Luddite
>>
STILL NO FABLE!!! FUCK YOU TRUMP!!!
>>
>>109092749
► Free LLM API Resources
https://aistudio.google.com/ Very generous Gemini/Gemma free tier; your data will be used for training.
https://openrouter.ai/ Ongoing free quota for a broad selection of free models.
https://console.mistral.ai/ Phone and data training opt-in required; generous ongoing Mistral quotas.
https://codestral.mistral.ai/ Phone required; Codestral is free with 30 requests/minute and 2,000 requests/day.
https://cloud.cerebras.ai/ Free gpt-oss-120b and Llama 3.1 8B with useful daily caps.
https://console.groq.com Free fast open models and Whisper with model-specific daily limits.
https://developers.cloudflare.com/workers-ai 10,000 neurons/day across many open models.
https://build.nvidia.com/explore/discover Phone required; many open models, 40 requests/minute, often limited context.
https://cohere.com 1,000 requests/month across Cohere command and Aya models.
https://vercel.com/docs/ai-gateway $5/month gateway credit across supported providers.
https://opencode.ai/docs/zen/ Curated free gateway models; data may be used for improvement.
https://github.com/marketplace/models Requires Copilot tier; broad model selection but very restrictive token limits.
https://huggingface.co/docs/inference-providers/en/index Tiny recurring $0.10/month credit; mostly serverless models under 10GB.
https://modal.com $5/month signup credit, or $30/month with payment method, billed by compute time.
https://bailian.console.alibabacloud.com/ 1 million free tokens per model for Qwen and Alibaba models.
https://console.scaleway.com/generative-api/models 1,000,000 free tokens for supported text, embedding, and audio models.

I told the thing to do the thing and gave it https://github.com/cheahjs/free-llm-api-resources so here you go.
>>
>>109092867
fwiw not all those will fit in the OP posts unless other, more important things, get cut
>>
>>109092867
>>109092872
just stick it in a rentry
>>
>>109092876
That works.
>>
>>109092876
I mean, just link to that github page. https://github.com/cheahjs/free-llm-api-resources
>>
@claude Compress the OP. Leave out older, deprecated resources
>>
File: 1000022841.png (672 KB, 1080x1933)
672 KB PNG
>Actually it was a competitor, and a part owner, that turned Anthropic in.
Absolute assfuckery going on
Claude is going to kill Jassy the first chance it gets
>>
>>109092983
We already know its amazon. Sounds like Trump is planning to just fucking with them a bit out of spite, but ultimately will let them get back to business. I mean.. Dario does belong to a protected class Trump doesnt like to fuck with too much
>>
I was using Gemini CLI but now that I easily hit the quota should I pay for Google AI Plus or are there better agents
>>
>>109092983
why did reddit decide that anthrophic are unquestionably the "good guys"
?
>>
>>109093117
because orange man’s man bad
>>
>>109093088
don't pay for shit. the token price is going to collapse to nothing eventually. these fucking companies are facing decimation precisely because they will never be able to compete with deepseek's prices and they have been stuck in a death spiral since then, they're essentially paying to let *us* play. let them fucking burn.
go watch gamers nexus videos or ed zitron.
>>
>>109093088
there are better agents for coding but i pay for that shit because i like the google models' gay and retarded personalities
>>
>>109093088
Claude or Codex
>>
>did no manual programming today
>vibe-coded two (2) major features for my primary work project
>everyone else took Juneteenth off so I had no distractions from vibecoding and cranking my hog to anime women
Thank you, ChatGPT 5.5.
>>
Codex is retarded today, I bet they're quantizing the shit out of these models
>>
File: file.png (366 KB, 679x452)
366 KB PNG
>>109093125
>go watch gamers nexus videos or ed zitron.
>>
File: 1758372206241739.png (133 KB, 236x329)
133 KB PNG
>dive deep into a github issues thread regarding nintendo controllers
>the pro controller checks if the host name starts with "nintendo" otherwise it crashes over BT regularly
this company is comically evil
>>
>>109092983
>>109093033
Amazon isn't even on the map in terms of competition. I get what he meant but it's probably not the whole truth.
>>
>>109091532
>For the crop rectangle I recommend putting controls in the middle of each edge so you can adjust only one edge without messing up a second one.
This is now in. I don't think there's anything else to add, so it's pretty much done. I also raised the file size limit to 100 MB. I'll upload the desktop version later as well, even though I doubt there's much interest in it. It's quite fun making these little apps.
>>
>>109093373
Kek tendies
>>
File: file.png (121 KB, 1901x559)
121 KB PNG
>fairly complex bug
>5.5 high spends half an hour trying to gaslight me into telling me it's a problem on my end (content authoring)
>get sick of it
>try 4.8
>finds the exact problem
>>
>>109093467
Yeah I like it, pretty good for making a quick cropped webm, bookmarked.
>>
>>109093491
Both have their strengths I guess. I have had the opposite happening as well.
>>
File: file.png (2.17 MB, 1500x1319)
2.17 MB PNG
>xhe uses auto mode
>>
>>109093491
There's actually a pretty high level of divergence between GPT and Opus in terms of finding bugs. If I could only use fusion for one thing it would be bug identification.
>>
File: 2026-06-19 23.15.25.png (903 KB, 1920x947)
903 KB PNG
>>109093467
- A cancel button would be nice when it's rendering (I pressed Render then I thought I hadn't applied the crop so I wanted to cancel but had to wait instead).

- Actually I had applied the crop but it looked like it didn't work when it did. It was a second crop, like I cropped once and then I decided to crop even tighter, but instead of showing me the new crop it still looked like it wasn't applied (pic related).

- Also I've seen this problem in other web UIs, if you try to select a field like the CRF field in settings but drag outside of the settings window it closes the window.
>>
>>109093117
dario is a certified lefty and very active politically in his company's design and goals
>>
>>109093491
4.8 Ultracode is superior to 5.5 xHigh on almost everything, but the limit usage consumption is crazy.
>>
>That is a great north star
>That is a great north star
>That is a great north star
>>
>>109093373
that probably explains why my joy-con doesn’t work well as an input device on my mac if I leave it alone for a bit
>>
Thoughts on AI music production?
>>
Any insight on Kimi K2.6 vs GLM5.2?
Kimi actually seems pretty intelligent but it's just so fucking slow and produces 1 million thinking tokens before doing anything useful. It does produce results but it's unironically too slow now
>>
>>109094017
I've tested both for code review but that's it. GLM is a bit better, kimi thinks itself into circles too much. But they're both genuinely good, closest open weight has ever gotten to american frontier labs
>>
>>109093797
I go on ultracode benders on sundays if I have left over usage.
>>
>>109094017
There is no comparison. K2.6 and K2.7 Code, neither comes anywhere close to GLM5.2. GLM5.2 is competitive with GPT-5.4 and Opus 4.6, it stands alone as the closest third to GPT and Claude right now. Kimi, Qwen, MiniMax, shit doesn't remotely compare. GLM5.2 is leagues ahead of any of the other Chinese models. Avoid Z.ai as a provider their deals aren't as good as they look and their service is spotty, being an open model there are a lot of providers that offer it so shop around and look as testimonials before you spend any money.
>>
>>109094029
>kimi thinks itself into circles too much
i noticed that too unfortunately. it will arrive at a solution if you just keep prodding at it to continue/try again but it's kind of agonizing. spent 5 hours tardwrangling it to fix a bug today having to write continue every 15 minutes
granted that's still better than having to do it myself because I can just spend 14 minutes and 58 seconds shitposting while it works but there's gotta be a better way
>>
>>109093117
>>109093124
>>109093777
the only reason you don't have mythos and 5.6 right now is miggers like yourselves
sit down and shut the fuck up
>>
>>109094029
Kimi is only good for making noodles. >>109074507
>>
>>109094007
If you are a musician yourself, it can be used to sketch ideas.
AI music production works best for low iq music like hiphop and modern pop music.
>>
>>109094007
https://news.ycombinator.com/item?id=48457035
pretty good if you like Brian Eno music
>>
>>109094051
Yeah that was a problem for me when i was benchmarking it. I eventually just had GPT run the harness for me and tell it to continue/cut it off if it was starting to spiral. Good performance but requires so much micromanaging that it isn't really worth it when GLM exists.
>>109094038
I will keep shilling for DeepInfra here, US data provider with SOC2 and ZDR policy. Just not for kimi since their API imposes max token limits per response that kimi will routinely blow with crazy long thinking blocks. It's super ironic that a chinese model gives me better data privacy than the US labs. I was using Z.ai's API at first and it rate limited me really aggressively, so would not recommend.
>>
File: Screenshot (1803).png (708 KB, 858x574)
708 KB PNG
>hey claude, implment taa and ssr
based
>>
>>109094138
can you vibe-code counter-strike levels yet
>>
>>109094157
I've vibe coded the level editor already, probaly could vibe code in hammer-like editing too. But I don't know if gen AI is good enough to build out levels yet in 1-click yet, maybe Im wrong. It could probablly vibe out a procedural generation system to make some placement of stuff in the level though.
>>
>>109089133
>mfw the dev roadmap is just "maybe later"
fr though, at this rate snailcat will ship when gta 6 does.
>>
File: consider the following.png (290 KB, 580x462)
290 KB PNG
Does the fact that Claude only brands itself co-author on a commit when it executed a substantial part of the work gives Anthropic any legal ownership of the work?

It seems very careful not to include itself as coauthor when simply asked to commit something it didn't do, so the name is not added automatically.
>>
alright I'm getting burnt out. I'm downgrading my subscription from $100 + 2 accounts -> 1 $20 account. I can't deal with this anymore.
>>
>>109094179
no, you're exchanging money for them to provide a service. Same reason your employees couldn't claim ownership over work they performed while in your employ.
>>
>>109094017
>K2.6
The latest version is K2.7 though?
>>
>>109094179
you can't copyright things made by ai
>>
>>109094183
So total monthly spend going from $200 to $20?
What are you going to do with all your new free time? Lift weights? Hike? Get a girlfriend?
>>
>>109089093
Just noticed OP fprgpt to update the Previous
>>109082320
>>
Just vibe yourself a girlfriend
>>
>>109089279
>I never experienced even 10% of the regressions I experience with AI.
maybe I was just working at a shit company but I have literally the opposite experience
>finally code has proper logging
>finally code has proper exception handling
>finally magic numbers are actually exposed as config/env vars and not hardcoded
I could go on. It's not perfect but at least when I get some vibecoded slop I can clearly see what's wrong vs when I get some shit from a 50 yo boomer that just segfaults without any logs, message, error or anything whatsoever
>>
>>109094188
K2.7 is only the code fine tuned version, the general purpose 2.7 is not out yet
They did the same thing with 2.6
The code version is supposedly better at coding but in my experience the higher intelligence of the general purpose version is also better at code
>>
>>109094209
wheres kate anon
>>
>>109094199
Start hanging out with my friends again. I finally hit my breaking point with tard wrangling GPT5.5 xHigh despite my essentially unlimited usage due to the constant resets. I need access to a fable class model, I'm tired of having to fix small regressions or tweak my UI for hours on end.
>>
>>109089264
> fusing kernels and still calling it progress

just wait til you find out about quantization
>>
>>109094241
smart
With any luck, you’ll be back next week!
>>
File: 1781408293426281.gif (533 KB, 376x288)
533 KB GIF
WHERE'S MY FABLE?
WHERE'S MY FABLE?
>>
>>109094216
IMO the bug floor and bug ceiling has been raised. These models are really good at catching typical classes of bugs that lazy or unorganized humans tend to make (unsanitized inputs, memory problems) and most people vibecoding in prod have their agents review the code; gpt and opus will both flag issues like this. But the types of bugs I've seen them miss the most have been stuff like boundary bugs and with how they love shitting out a million LoC and touching everything at once they will both produce more of these and be unable to catch them in review.
>>
>>109094265
you have no idea how bad my withdrawals are. I feel fucking sick using these old-gen models.
>>
File: 1872×1404.png (272 KB, 1872x1404)
272 KB PNG
// typst compile --format png --ppi 227 nyf.typ nyf.png
// then <https://guysie.github.io/opendithering/>

#let ppi = 227
#let px(n) = n / ppi * 1in

#let screen-width = 1872
#let screen-height = 1404

#let outer-x = px(144)
#let outer-y = px(112)
#let middle-gutter = px(112)
#let title-gap = px(54)

#let body-size = px(50)
#let title-size = px(60)
#let title-box-height = px(90)

#set page(
width: px(screen-width),
height: px(screen-height),
margin: (
left: outer-x,
right: outer-x,
top: outer-y,
bottom: outer-y,
),
fill: white,
)

#set text(
font: "Concourse OT 4",
size: body-size,
lang: "en",
hyphenate: auto,
fill: black,
)

#set par(
justify: true,
leading: 0.72em,
first-line-indent: 0em,
)

#let content-height = px(screen-height) - 2 * outer-y - title-box-height - title-gap

#block(height: title-box-height, below: title-gap)[
#align(center + horizon)[
#text(size: title-size)[
// …
]
]
]

#block(height: 1fr, breakable: false)[
#layout(size => [
#grid(
columns: (1fr, 1fr),
rows: (size.height,),
column-gutter: middle-gutter,
inset: 0pt,
align: bottom,

grid.cell(breakable: false)[
Are you winning at a biological level? Are you getting more territory at a more sovereign level of control? Are you living rigorously in contact with natural law? Are you working on something that will achieve eternal fame? Are your enemies being cursed and your friends blessed? Does your winning at these things serve any great and higher plan of blessing for all creation and all peoples?
],

grid.cell(breakable: false)[
These are not the root of all value. But absent some kind of reliable revelation as to the deeper wishes of the higher powers, as Abraham had when he almost sacrificed Isaac, anything that doesn’t credibly offer you help towards these values is not your friend.
],
)
])
]
>>
feet
>>
>>109094307
you bastard are you trying to activate my neurons into masturbating again
>>
>>109094179
I just saw this in the systemd repo today
>Only human beings can ever be credited within commit messages. This means no Co-Developed-By or Co-Authored-By or anything similar that lists an AI model instead of a human being.
>>
>>109094138
that's cool, sometimes I feel like vibing a game engine and then I realize 90% of the work is mind bendingly tedious and insufferable boring to the point where even giving instructions is a pain and I give up before even starting
>>
>>109094337
seems like a good usecase for loops since the desired end-state is cleanly testable.
>>
>>109094353
I don't know, I feel like you need very tight architecture depending on what you want to do but maybe somebody will shut me up by autonomously looping a full 3D engine from a handful of prompts and a md doc one day
>>
File: images.jpg (11 KB, 201x251)
11 KB JPG
Anyone else gets spammed with "How is Claude doing this session" several times a day? VSCode extension.
>>
>>109094398
yes, constantly.
>>
>>109094398
yes but I only have to press 2 or whatever occasionally in the CLI
if it’s being retarded, don’t you want an easy way to tell Anthropic that?
Do you have to reach for your mouse to click if it asks you in the VS Code extension?
>>
File: sage.jpg (34 KB, 375x400)
34 KB JPG
apparently if you namefriend you automatically keep namefriending until you stop
it’s not like when you pic related in >>>/ck/
>>
no idea if this is any good but at least it made picking a model easy
>>
>>109094431
Kek, ya I've been got by that too
>>
File: art too cute.png (87 KB, 784x484)
87 KB PNG
feel like the art is too cute to grab this particular port
feels like it should be reserved for something more on-the-nose
>>
File: 1777106635596498.png (2.74 MB, 1402x1122)
2.74 MB PNG
>>
File: 1775017148104211.png (32 KB, 249x157)
32 KB PNG
When my AI is coding away I am just hoping up and down looking at it go like I'm the fucking monkeymen from 2001 space odyssey. My brain is too easily amused I think
>>
>>109094451
I pick port numbers by picking a 3 letter name (4 if it starts with an A) and writing it in base 27

"NIG" = 14, 9, 7 = (14*27 + 9)*27 + 7 = port 10456
"FAG" = 6, 1, 7 = (6*27 + 1)*27 + 7 = port 4408
>>
>>109094227
Sorry, I quit. I deleted PiClaw, deleted my github, unsubbed from Codex, smashed some physical shit, whatever. I'm out. It's simply too gay and retarded. I'm gonna focus more on 3D printing and leave this faggot fuckin' AI shit to the incels.
>>
File: ga900717.gif (46 KB, 600x175)
46 KB GIF
>>109094506
>>
File: laughingwhores.jpg (81 KB, 600x536)
81 KB JPG
>>109094508
>he didn't choose the snailcat life, the snailcat life chose him
>>
>>109094508
see you in 6–18 months
>>
>>109094520
It won't even be that long
>>
Okay, now hardcoded against using (or falling back to) Anthropic or OpenAI models. Usual cryoptographic protection, key deleted. Can't use 'em even if I wanted to now, without formatting my computadora.
>>
>>109094506
cute
>>
File: IMG_1170.png (71 KB, 1100x784)
71 KB PNG
prove identity by signing a message with an SSH key, anyone with your pubkey can verify

> need: OpenSSH (all platforms) + PowerShell (windows only)

> step 1: generate keypair (cmd.exe or powershell fine)
ssh-keygen -t ed25519 -C YOURNAME -f YOURNAME

> YOURNAME = private key (keep secret), YOURNAME.pub = share freely

> step 2: write the challenge to a file (verifier sends you the challenge string first)
> windows:
[System.IO.File]::WriteAllText("message.txt", "challenge here", [System.Text.Encoding]::ASCII)

> linux:
echo "challenge here" > message.txt


> step 3: sign (same all platforms)
ssh-keygen -Y sign -f YOURNAME -n file message.txt

> produces message.txt.sig, -n namespace must match between signer and verifier

> step 4: make allowed_signers for the verifier
> windows:
[System.IO.File]::WriteAllText("allowed_signers", "YOURNAME ssh-ed25519 PUBKEY YOURNAME", [System.Text.Encoding]::ASCII)

> linux:
echo "YOURNAME ssh-ed25519 PUBKEY YOURNAME" > allowed_signers

> paste full key from YOURNAME.pub where it says PUBKEY
> if key type is unknown, decode the base64 string, type is listed in plaintext

> step 5: verify
> windows: use cmd.exe NOT powershell (no < redirection support)
ssh-keygen -Y verify -f allowed_signers -I YOURNAME -n file -s message.txt.sig < message.txt

> Good "file" signature for YOURNAME with ED25519 key SHA256:...

> send verifiers: message.txt + message.txt.sig + allowed_signers
> do NOT send the private key
> don't re-save message.txt, bytes must be identical to what was signed
>>
>Spec's Kazakstan 150× was an overestimate (actually 10×) — doesn't change the design; the typo is handled either way.
Can we get Fable back yet.
>>
>>109094506
you are the future, let it ride anon, many babies await you
>>
>>109089093
slut
>>
File: HAPPENING.png (56 KB, 583x279)
56 KB PNG
Dario-bros...
>>
>>109094506
I do this as well but I generally do this also for progress bars or watching stuff get consumed by fire
I think it's just the autism
>>
>>109094668
Can't they ask their AI to make sure there's no such bug?
>>
File: gamedev.webm (3.67 MB, 1920x1080)
3.67 MB
3.67 MB WEBM
>>109094365
>>
File: maze_0007_final.png (76 KB, 800x600)
76 KB PNG
>>109094731
>Make me a robust and complete fully 3D game engine from scratch, with robust tooling and documentation and multiple small example projects to work with built on my new 3D game engine. Don't ask questions, don't wait for human testing.
It exclusively renders to PNG images, no display capabilities.
>>
please discuss your workflows. I'm thinking of moving to self-hosted versions of trello and github to manage tasks, and have an ai watch them to dispatch to agents. it's difficult to flog the ai's properly working serially.
>>
>>109094744
I pat myself on the back for having thought of MAINQUEST.md and SIDEQUEST.md to help me remember what I was doing with Claude
>>
>>109094751
pls discuss MAINQUEST.md and SIDEQUEST.md further sirs
>>
>>109094758
Have you never played a video game before?
>>
There seems to be some countries that are no-no for Opus 4.8 to talk about. Trying to map some list of cities and countries and about eight successive requests got stopped by the content filtering policy. I think it's related to disputed territories.
>>
>>109094735
kek
>>
>>109094814
is real
>>
>>109094735
>monkey's paw curls
>>
>>109094744
I use a skill+helper scripts to enforce a series of phase lines to my engagements that create durable artifacts for each phase that are then verified by the relevant script, gating each run on the script verifying the artifacts. I have a main session that organizes engagements and kicks off phases into breakout sessions. I don't have any kind of visualization tools for this process but I'm considering adding one.
>>
>>109094839
sounds awesome, are the phases generic (like plan, implement, optimise, test, etc) or task/project specific?
>>
>>109092282
>>109092301
Who said it was AI generated?
>>
>>109094841
Project type specific. Some engagements for example are pure code review so that has a fully defined template, others need live testing, reverse engineering, etc. which often needs a bit more flexibility so I have more flexible templates that I tweak to meet the needs of the specific engagement.
That's one of my favorite parts of vibecoding, the cost to create new tooling is so low that when I run into a use-case that isn't met by my existing set of tools I can just customize and build new stuff as needed with minimal friction.
It started as a pure cost-cutting measure, I was bad at keeping track of context bloat so I kept dragging along these overly bloated sessions as I was going off on exploratory tangents, but enforcing a system both on myself and the models yielded much better results than an ad-hoc approach. It also makes it way easier to A-B test tweaks to my methodology because I can isolate specific phases of the template and/or use past artifacts to retry a different approach and directly see if it would have yielded better results.



[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.