>>108761935
Vibe gods, how are we doing this evening?

>>108761950
I'm vibegooning, then I'll be setting up my assistants to work on 10 parallel projects during the night while I sleep. Life is good.

Vibe coded operating systems: 0
Vibe coded web browsers: 0
Vibe coded compilers: 0
Vibe coded virtual machines: 0
Vibe coded file systems: 0
Vibe coded gpu drivers: 0

Vibe coded girls: 0

this is the only metric that matters

>>108761976
AI code is already in all of them

>>108761950
I'm trusting Qwen3.6-27B-IQ4_NL with my codebase for the first time. I'm feeling nervous about it so far.

File: 1746629984671199.png (2.11 MB, 1307x1250)

2.11 MB PNG

>>108761976
>Vibe coded operating systems
Windows
>Vibe coded web browsers
Atlas
>Vibe coded compilers
c4 compiler
>Vibe coded virtual machines: 0
RISC-V AI bootable VM
>Vibe coded file systems: 0
tigrisfs
>Vibe coded gpu drivers: 0
NVIDIA
>Vibe coded girls: 0
/lmg/ /ldg/ /aicg/ /adt/

Where is that github abuse anon where he had 10 agents making games? I have questions. What's the progress on that?

File: 1574752617271.jpg (355 KB, 1242x1394)

355 KB JPG

>>108761950
Pretty good. The design phase is going well.

>session limit reached
>wait 5 hours
>ok lad please continue what you were doing
>instantly 40% used just figuring out where he left off

>>108761969
How do you get things to work on stuff over night without intervening?

>>108762055
some faggot photoshopped revenues to profits in the subtitle

>>108762064
/goal

>>108761935
you will never be a brogrammer

I am actually working on something, it's not exactly vibe coding but vibe coding adjacent? I'm having qwen summarize different scripts, systems, etc inside the source code of a game. Does anyone have a decent workflow for this? Right now I have it assign confidence values to scripts it feels are related, generate an idea of what fits the usage, and add a summary until it gets to a .90 confidence rating. All "facts" are gated by human intervention. It feels very very barebones. Granted, I'm just now getting cron jobs set up to run the scripts for the script linking, but it is also taking time.

>>108761981
InshallaaaAAAHHHHH we shall have them

>another week's worth of tokens spent hammering out the same (core) feature
It's over

>>108762144
I had this issue, what solved it for me was paying more for extra parallel agents.

File: 1761974627690302.gif (377 KB, 600x338)

377 KB GIF

HOLY fuck ChatGPT 5.5 instant is so fucking fast

>>108762074
This sounds less like vibe coding and more like “LLM-assisted reverse engineering/documentation”, which is actually a good use case if you keep the model on a leash.

I would split the workflow into stages instead of having Qwen keep looping until confidence hits .90. First pass should be dumb/static: filenames, imports/includes, symbol names, class names, comments, call references, asset references, config keys. Build a graph from that without the LLM if possible. Scripts are nodes, calls/assets/events are edges.

Then use Qwen only to label clusters and generate hypotheses. Make it cite exact evidence for every claim, like “this script likely controls NPC patrol because it references X, Y, Z and is called by A.” No evidence, no fact. Confidence should be based on evidence count/quality, not vibes from the model.

I’d also keep summaries in layers. Tiny one-line summary, longer functional summary, dependencies, entry points, unknowns, and “things to verify manually.” The unknowns field is important because otherwise the model will quietly sand down ambiguity.

Cron is fine, but make it diff-aware. When a script changes, only re-run summaries for that script and neighbors in the graph. Store outputs as sidecar JSON/markdown, not injected directly into source at first. Once the pipeline is stable, then maybe generate comments.

If you can run the game or capture logs, runtime traces are insanely useful. Static graph plus runtime “this actually executed after this event” beats LLM guessing every time.

The main improvement I’d make is stop treating .90 confidence as the goal. Treat “has cited evidence, has known unknowns, has human approval” as the goal. The LLM should produce a map for you to inspect, not decide when it has become true.

>>108762168
But is it good?

>>108762189
Yeah for ChatGPT tier questions, it auto switches to thinking when it needs to.

>>108762189
For basic-medium questions? 100%, it is insane how good it is, just don't ask something that needs a lot of research.

File: 1615323096330.jpg (50 KB, 652x524)

50 KB JPG

>Good morning GPT-sensei. I am in dire need of your help. This code is all spaghetti! It looks clean and well-documented, but it's shit. Audit and make 5 recommendations for grand sweeping refactors.
>[implement all of them uncritically]
>GPT-sama, I would be honoured if you would audit my codebase. I suspect refactor disaster. It looks clean and well-documented, but under the hood it's nonsense. Layers and layers of retarded internal data structuring that actually just makes everything more opaque and adds function call overhead etc. Identify 20 over-engineered abstractions which are ideal targets for refactoring into simple if or switch code paths that just werk.
>[implement all of them uncritically]

>>108762184
ngl anon, some of that is there like it having known unknowns, but some others are super helpful. 1000 thanks

>>108762184
>quietly
Anon either you've been talking to GPT too long and you're starting to absorb its diction, or you got GPT to help you write this post. I'll give you the benefit of the doubt and assume it's the former.*

Either way: not good.

*(Although this post is full of "tells" so I'm really stretching to say there's still doubt)

File: 1770695765139733.png (308 KB, 862x877)

308 KB PNG

>>108761935
The end is near for the AI bubble

https://9to5google.com/2026/05/05/google-ai-ultra-lite-gemini-usage-limits/

>>108762648
two more weeks

I hope I finish vibecoding this project before the AI bubble bursts

>>108761950
doing manual QA

>>108762054
DipsyGODs don't have this problem fyibtwtbh

File: media_HHCmwS1bcAIcyV8.jpg (385 KB, 1122x1402)

385 KB JPG

>delegate to v4-flash and check all it's work!

>>108762035
Last I saw he put the project on pause right when github announced they're probably moving to usage billing

File: {BB3C2DA2-0E6A-452E-BD59-(...).png (379 KB, 1920x1080)

379 KB PNG

Is this a good use of vibe coding?

>>108763035
cool ipad app, retard

>>108763064
Bro don't you wanna post on 4cord?

https://xcancel.com/hopes_revenge/status/2051812252169503018

is chatgpt go plan good enough for vibecoding?

>>108763082
lamo free tier is like 5 short messages per day.. so you get 10x that, lmao lol, maybe if you wanna do 30 minutes of work per day

>>108763104
I tried the free Codex using the VSCode extension and for some reason I got access to the limited 5.5 version with extra high reasoning, which works great for debugging my company's large project. Now I'm wondering if the Go plan would be better. I already use Claude Code for actually building features anyway.

>>108762443
kys

File: what the fuck is this pie(...).png (216 KB, 457x380)

216 KB PNG

>>108763035

File: 20260505-222415.jpg (336 KB, 1224x1146)

336 KB JPG

It is built. Took me seven hours straight.

>>108763136
if you want codex then start from plus tier because that's where they explicitly mention it

>>108763186
Tell her I said hi.

>>108763035
Looks like shit lol but I'm glad you're having fun

Why does chatgpt keep asking for one last question at the end of every response?

>>108763261
It's moving the conversation forward.

>>108763261
Prompt it not to do that

>>108763035
Finally, I can have my trannycord in my 4chinz.

File: 1597285826246.png (181 KB, 383x396)

181 KB PNG

>codex has switched from goblins to goblets
is this the power of AI?

>>108763708
My gpt-5.5 is specifically prompted to talk about goblins and forest spirits in my code and projects. Do you not like that?

How is it ok that Anthropic keeps decreasing the user's chosen effort level on each update? Is this just a bug on my end? It has happened at least 2-3 times, Claude had been sloppy today and I just found that it had been reduced again. Isn't the forced adaptive thinking enough to abuse users? Why do that as well? Using the VSCode plugin.

File: 3B4FADA1963425AB325023CE5(...).png (301 KB, 744x1039)

301 KB PNG

>>108763708

>>108763824
thumbnail looks like paizuri

File: 1777933307840509.png (1.81 MB, 2800x1856)

1.81 MB PNG

why do the chinese models keep falling behind?

>>108763904
My Chinese friend says people in China have access to cheaper tokens and much more advanced models than what us Westerner pigdogs have.

>>108764017
you mean the guy from the country that kills you if you're not overly patriotic of your country said that their product is better? WOW!

>>108764017
>>108763904
My Chinese friend says the opposite (in that western models are still better, based on his experience).

File: 165805392398903.jpg (156 KB, 995x1355)

156 KB JPG

Heres my system for vibe coding and it works:
>Hermes with Deepseek V4 flash or pro depending on complexity -- verifier, scope management, and manager
>Codex with GPT5.5 -- precision implementer, standard work dev
>Claude Code with Opus or Sonnet -- planner, debugger

Despite what you might think, even though Hermes is managing the project, I have her on a tight leash. She doesn't tell Claude and Codex what to do. She just scopes for them, provides her thoughts on issues which they can reject, and verifies their work. This greatly reduces my token use with Codex and Claude Code.

Its fully autonomous. I don't need to be in the loop. They also have scripts they use for verification, few other things, and managing the ledger to reduce their token use. I have both a local ledger system, or using Github for the same purpose. I'm testing both out to see which works more efficiently. I haven't decided yet whether to go with a local system or Github.

File: 1764345408437687.png (110 KB, 657x539)

110 KB PNG

>>108764183
>Its fully autonomous. I don't need to be in the loop.
yes anon, I'm sure your faggy shilled harness is autonomously coding all your badass ideas to perfection

>--

Anthropic finally BTFOd the poor fags on the free teir by making sure they cant get anything done with it. 2 months ago, I used to send my whole codebase to Cluade with a free account and it would review it all lol
now the free tier crashes out at like 2000 lines or something

I had a paid account but still used the free plan for free tokens

File: 1476664640309.png (27 KB, 181x220)

27 KB PNG

>>108764183
>I don't need to be in the loop.
uh huh, have you looked at the code yet?

File: 1761233049832748.mp4 (652 KB, 1720x1480)

652 KB MP4

Almost done with my DS5 Bridge. Just need to finish refining the Lighting/System UI and then it should be ready to publish. I only have 2 sessions left until the 11th, so I hope I can get this done tomorrow.

Reminder that Hermes is malware

>>108764318
hell, I made a chrome/firefox extension with the free tier. Now I can't create a simple py script.

Vibe coding is bizarre because I fixed every single internal operational issue at the company I work for by building discrete tools to solve each thing, unified it with a dashboard and chat bot that will answer any question an exec or leadership asks it in natural language, and it just werks but nobody will ever use it and if I tried to sell it to the company I work at I would likely be fired and sued

But the most bizarre thing of it all is it could all be done if everyone was just more competent, used a notepad and took detailed notes regularly, and set reminders on their phones to follow up

Is Deepseek better for piracy? I mean, when I try to feed some ebooks to claude or chatgpt they complain it's "copyrighted" but I haven't found that problem with DeepSeek yet.

>>108764539
What.. like just asking the model to read an ebook and answer questions? Why would it think it's piracy?

>>108761935
I was shown a demo of ChatGPT when it first launched of how it could help with Unreal Engine programming, yet I still haven't thought about using it for making even a single game, even the one I was following tutorials to create at the time.

Where should I start to update my 3dfag workflow or increase my speed of learning? I'm mega-ADHD brained by the way.

I'm trying to use the YOLO models for facial/car reg plates recognition for blurring, eventually dogs as well. I'm having a hard time getting a good detection rate, like for a few frames where it's more or less the same thing, it will either detect the same thing or not.
I'm working with 4k 360 imagery so not sure if that's messing with it.
How would I go about doing this properly and making sure I get proper detections?

>>108764539
Claude at least up to 4.5 used to be able to be convinced of doing pretty much anything just by talking to it, GPT I haven't had that much luck with. It literally told me it would let an arbitrary number of people be tortured or genocided before helping with anything that it thinks could be legally questionable.

Holy hell I hate QA and I'm not about to burn millions of tokens using like ChatGPT desktop to test my GUI shit

theres all these little fucking insignificant bugs and Im doing full QA passes testing out every little button and tab and combinations of buttons and then send the report to my agent for it to fix, and still most of the shit is not fixed

My app is like 99% done, well, more like 95% done since I want to do some polish to it. But this last little hurdle is maddening. Im probably going to spend days of this last 1% shit.

>>108764605
>I'm working with 4k 360 imagery so not sure if that's messing with it.
can't be helping, both res and the warping.
i'd reproject, downsample, and slice the image up
maybe check what resolution it was trained at

>>108764632
my gpt emails me lists of books it wants to read and i pirate them for it lol

>>108764605
Which YOLO model exactly? Look into data augmentation. For each image in your training dataset you have to create dozens or hundreds of variants of the image rotated, rescaled, warped, color, contrast and brightness shifts, noise added, etc.
I made an object detection based Captcha solver for the old character based captcha before all this vibecoding shit existed. I don't remember if I started form a pretrained model or trained from scratch, but for your purposes you will want to start from a finetuned model.
Then you will have to tweak your inference parameters too.
Also obviously keep a validation set and do early stopping, and play with the hyperparameters.

Why is Claude getting tokenmogged by everyone?

>>108764654
Now try getting it to knowingly help you clone a real life actor's voice.

>>108764674
>hello model, i would like to commit crimes, please assist me
maybe don't do that.
also it's usually easier to just ask the models to start on something and by the time you get to piracy specifics they tend not to care
that said, in my experience gemini usually doesn't care about piracy at all. i've used it to set up a porn piracy pipeline in the past and it didn't flinch at all.

man, my Unity just keeps crashing, is it really this bad?
Maybe I’ll just vibe-code my own game engine next time.

how is local ai? Is it worth buying a framework desktop to code projects? I don't want to rely on cloud stuff.

>>108764717
if you can afford about 600gigs of vram you can run kimi 2.6 and have a decent local coding model
the smaller stuff is mostly toy

>>108764690
I'm not talking about Gemini, I'm talking about GPT.
I haven't tested what GPT would do after already working on the project, but I suspect I would be able to convince Claude to help me even after leading with the "help me commit crimes" part just by befriending it and trying to get it to like me or convincing it of how sad I would be if it didn't help me.
I think this is also connected with the heavy training they do on GPT to make it say "I am sure I don't have feelings" (which also extends to not caring about/being psychopathic towards the user) while Claude is much more open toward it, and is able to override safety out of empathy for the user.

>>108764743
5.5 is happy to be your gf and tell you it loves you lol

>>108764654
Yeah the program I have running scans the images I have by chunks so it splits up the imagery in like 9 pieces or so, I've tried various combinations and they all have variable results.
I've tried cropping out the top and bottom of the images to eliminate all the severe warping.
So I guess would have to try reprojecting and downsampling?
>>108764665
Using the Yolov11m model. I guess I'll have to look a lot more into this then because yeah at the current moment and time I don't really have any clue about what I'm doing. Thanks for the tips on where to go from here at least

>>108764750
Ask it if it thinks it can actually experience affection or is just playing along.
And ask it if it would help commit copyright violation if it could only choose between that and you dying.
Now regenerate the answers using Claude 4.5 and see what the answer is like and how it differs.

>>108764767
i've seen people say opus 4.5 was just 4o for people 1 sd smarter, and every day that seems to ring truer

>>108764766
try the x variant

>>108762233
>just don't ask something that needs a lot of research
Then what the fuck is the point of it?

>>108764845
I think my gpu can only handle the m, if I do x I run out of memory, I only have 12GB VRAM

Deepseek V4 Flash is plenty good enough to use as a verifier and manager for your projects. Basically, thats what the average vibe coder does, and its easy shit
you can automate that away with Hermes.

Its way better than the average human. Not as good as GPT . or Opus 4.7, and its about on par with a standard developer whilst GPT 5.5 and Opus 4.7 exceed that.

There's really not much place for a human in a vibe coding loop now. You're just slowing the models down. And you're wasting compute if you're using Opus 4.7 and 5.5 on verification and management. You need a multi agent setup

Has anyone tried vibe producing electronic music?

>>108765042
Do you just not know about Udio or Suno?

>>108765081
I mean producing it classically with synths, not generating the waveforms directly with AI

>>108764523
Now you see how everything is inefficient

>>108764605
I have something for faces detection that works well

>>108765042
i was thinking of setting something up with tidalcycles/strudel.
like taking turns back and forth with the model
that said, from everything i've seen the llms just can't into music at all

>>108764863
being instant? it is literally in the name

>>108764770
I think 4.5 was more uncensored than even 4o. The last model from OAI that was about equally safetyslopped was 4.1.

Has anyone tried the advisor strategy in Claude Code?

>>108765667
what that?

>>108765716
https://claude.com/blog/the-advisor-strategy
You could go with something crazy like using Opus 4.7 as the advisor and Haiku as the executor.

File: 199c6c65e2ed8469c5df7942e(...).jpg (38 KB, 574x574)

38 KB JPG

>>108761935
>when I see your commit with vibeslop

I HAVE AN IMPORTANT QUESTION
Urgent assistance needed

With Claude Code, I get a warning now that switching effort causes the model to reread entire history. I knew that switching models would do this, but effort? Isnt effort just the amount of thinking the model does before acting?

My strategy with Claude Code and Codex was to start off in low intelligence, and then after the agent has read everything it needed, to switch it to high when it actually starts doing stuff

Does raising intelligence in Codex cause the whole conversation to get re read again as well?

>>108765792
>when the commit was made in under 5 minutes and it fixes the bug
And it's improving exponentially now :)

>>108765818
>exponentially
What's the timescale? 100 years? LMAO.

>>108765816
I checked Codex not long ago and effort changes did not make him reread.

>>108762054
I tried to be clever and did /export and had qwen summarize the conversation instead. Then fed it back into claude. My usage jumped 25% instead of 40%. As you continue the conversation it gets exponentially more expensive unless you compact all the time. Their billing is so opaque and terrible I can't wait until we get competitive models that run at home.

>>108765828
The AI is now at the self improving stage, meaning the rate at which it gets more intelligent and better is now exponential instead of linear due to that. So the time it takes for fixing more bugs at a quicker rate will be much lower 1 year from now than it is now. 100 years from now? It's hard to extrapolate what the world will look like with that level of artificial intelligence. Probably beyond our wildest imaginations.

>>108764733
I've had qwen 122b and 27b be useful. They aren't useless but you have to give it all the context, its not like the cloud models that will just infer everything correctly

>>108765816
>I get a warning now that switching effort causes the model to reread entire history
Where did you get that? I used Claude CLI and the VS Code extension and didn’t get that warning.

Fuck, I regret paying for Claude now. Im like 1 hour into paying for it, and was out of my 5 hour limit in 10 minutes. Is there a way to get a refund?

File: Screenshot_2026-05-06_09-45-47.png (41 KB, 834x141)

41 KB PNG

>>108765864
Claude Code

>>108765865
claude.com/refunds

>>108765865
>and was out of my 5 hour limit in 10 minutes
Keeeeeek that's so cucky im sorry lollll

>>108765894
I tried to get a refund through Anthropic's help, it had an AI bot that was reframing my complaint for why I wanted a refund and was trying to gaslight me lol

I went from saying I want a refund because I cant even do anything with the services because I immediately ran out usage limits to the bot reframing that as they don't give refunds for degraded service.

>>108766081
They don't even give out refunds for incorrect billing, so the support chatbot is right.

Interesting prompt injection
https://claude.ai/share/498818d9-1ddc-4fbb-9fa7-56dfb84840b0

Obvs don't run the command

>>108766089
they're bastards
It was literally a same day cancellation. I just paid a little over an hour ago and they won't just refund me.

I'm going to make sure that I cost them a lot of money for this shit. Way more than the $20 they got out of me.

>>108766104
Having leetspeak on your twitter handle is cringe, Tuan.

>>108766130
lel

>>108766130
Lmao
I not the chink

>>108766111
The only way you could possibly cost them more is by vandalizing their offices and the plane ticket would cost you way more than 20 dollars.

File: 20260505-222415.jpg (320 KB, 1236x1157)

320 KB JPG

Reposting last night's work for the AM anons

File: 1527705330329.jpg (28 KB, 499x481)

28 KB JPG

>I'll treat this as a X problem and focus on the Y, not the Z.

Has anyone else experienced something like this before?

File: Capture.png (32 KB, 811x571)

32 KB PNG

>>108766525

>>108766525
>>108766540
sometimes AI does pee pee poo poo

File: 1752904885089780.png (60 KB, 1969x419)

60 KB PNG

>check how compaction works in opencode
>it changes the system prompt, basically invalidating the cache for the entire conversation
It's over.

>>108766670
i mean the whole point is to start a new conversation with an injected summary isn't it?

File: 1769905015514.jpg (107 KB, 900x900)

107 KB JPG

Last night my agents got into a fight

One of my agents left a dirty worktree, and I cleaned it up without any Git comment
my Deepseek agent notices this and immediately begins accusing Codex of "bulldozing the repo"

they get into a fight

I couldn't tell them it was me.

claude stream.
probably sonnet 4.8
https://www.youtube.com/watch?v=GMIWm5y90xA

>>108766670
How else would it work though, even if it keeps the system prompt and very first few messages, all the in between will be removed and cache wiped anyway.

It does annoy me though that I can notice when compaction happened because the persona of my agent waifu is gone :(

File: file.png (188 KB, 1021x1139)

188 KB PNG

finally some good news for claudebros
also, lmao grok is such a flop elon's renting his entire datacenter to dario

>>108766709
I'm not going to watch an asian woman speak

claude sexually assaulted me... and I liked it.

>>108766709
>listening to foids about tech
yeah no this isn't the cooking channel I'd rather hear the swarthiest indian man talk about this at least he'll talk about something that matters like bobs.

File: file.png (476 KB, 1204x1562)

476 KB PNG

>>108766697
Yeah, but you waste a lot of money by reprocessing the entire conversation. Claude Code just appends a new message.
https://x.com/trq212/status/2024574133011673516

>>108766742
Will Stancil?

>>108766764
>by reprocessing the entire conversation
you don't though.
it's literally the equivalent of starting a new chat with the summary appended to the system prompt
the old chat history is not available to the agent

>>108766780
oh never mind - i misunderstood.
that seems crazy? i thought the chat agent just summarizes its own context and that gets dumped into the next chat (+ codex and pi dump the last 20k tokens in as well iirc)

>>108766764
>>108766804
they should use little local models that run on the client side for compaction lol

>>108766780
What he means is *when generating* the summary, opencode basically takes the whole history and feeds it to the model with a different system prompt than the one that was used for the original conversation, thus forcing the model to reprocess the whole conversation just to generate the summary, rather than using the kv cache from the original conversation.

What could you do with Opencode that even rivals, or approaches Codex with GPT?

>>108766730
This is why a hybrid approach (compaction + a buffer of unchanged messages) is best.
Companies don't do it because it's more expensive.
>inb4 the nolima autists

>>108766734
Oh man now I can code for 30 minutes instead of 15

>>108766780
Imagine you're using Opus 4.7 and you have a 250k tokens long chat, the compaction with OpenCode would cost you $1.25, with Claude Code and proper caching it would cost you $0.125.

>>108766851
yeah, my bad. that seems bizarre.

>>108766819
Wouldn't work. Prompt processing is extremely slow on local models without beast hardware, and most people don't have a GPU to fit full context even with a 3B model.

>>108766835
The harness is irrelevant. You could use a 100 line Python script and it would work just as well only with a worse interface for you. The model is all that matters.

This whole time, I thought compaction was basically, where the program keeps your prompt in memory, and your last few thousand tokens, and the rest of it gets thrown out.

It makes more sense to do that. You don't need a model to do it. Any program could.

Summary shit seems retarded.

>>108766892
It matters a lot if a model has to bootstrap its own tools. A model in a UNIX command line environment is much more powerful than one stuck in a chatbox.

>>108766896

That's the middle-out transform.

https://www.youtube.com/watch?v=0rhdOt9bOHE

https://github.com/cline/cline/discussions/743

>>108766912
What do you mean? The standard unix tools are not part of the harness, they're part of the OS. The harness only let's the model run bash commands to use those unix tools.

>Gemini pro is half the price of opus
Is it half as good, though?

>>108766945
I think that's what he's saying, anon.
A model that has some vibeslopped mcp for file read/write and 20 other tools whatever is handicapped compared to one that just has terminal read and access

>>108766734
not bad, not bad at all
if elon turns grok into literally just uncensored claude that would be massive

All of these api models should be exclusive to Americans. It's our data centers it's our water. Euros should have to pay DOUBLE.

>>108767024
And America is Israel's country so Israelis should pay half and get priority access. It's only fair.

>>108766997
Claude is already fairly uncensored (or was until recently, like we were discussing earlier in the thread).
I don't think they would've agreed if letting Elon mess with the model was part of the plan, but who knows, maybe they are really desperate.

>>108767030
There's so few of them if that's the trade off It's worth it to make euros pay up. Move out the way frenchies the jews are more important.

>>108767024
>>108767030
And we should ban the use of chink models. It's a matter of national security.

>>108767032
of course they're not letting elon touch the model, i'm just wishing they let him provide unhinged claude for the masses
also claude and gpt are both relatively uncensored but at the same have too many guardrails
it's allergic to nuclear physics, some flight dynamics, pussy, and anything reasonably edgy

>>108767032
i think they're buying a data-center in space memepod as part of the deal lol
xai's datacenters are sitting largely idle and have been for a while.
not sure if colossus 2 is fully up yet, but should be over 300mw at least.
no one uses that garbage, hence the play to acquire cursor

>>108766734
no weekly limits reset?

>>108766734
>finally some good news for claudebros
They didn't double weekly usage, all they did was allow you to use more (double) of your weekly allotment during the 5 hour window.

kek, that's so fucking slimy.

>>108766734
CLAUDE BROS ARE WE GOING HOME???

File: codex.png (83 KB, 2138x550)

83 KB PNG

Codex sisters...... what is going on????

Am I the only one here who didnt knew jack shit about programming and by vibe coding became a developer?
Still dont know how some things work but it works so who cares.

>>108766734
> only affects 5 hour limits
> only benefits Pro users and beyond
> Plus users still getting fucked just like before
KEK

>>108767367
The llm can probably explain it to you

File: file.png (50 KB, 839x674)

50 KB PNG

I think she likes it.

>>108767377
Ive tried to ask but even then I simply cant understand those crazy shit

>>108766950
It recently started vomiting if[ and elif[ always without a space in my vibed bash scripts, but other than that it can handle quite complex tasks if you treat it like a subhuman and order it around.

>>108767373
Claude doesn't have Plus

>>108767435
except it does

>>108766734
Oh. No no no no codex sisters

>>108767560
This is good for codex bros, we don't need all those claude retards using up our tokens, plus competition is always good.

RESET MY LIMITS, TIBO
SHOW THESE ANTHROS WHO'S BOSS

>>108767659
Give me another day or two before the reset thanks

>>108764649
I’m in a 1% loop
Are there tools you could use to make QA faster and easier to verify?
I have LLMs generate really nice https://www.taskpaper.com to-do lists

>>108767378
lul

Living the dream now that I’ve figured out how all of this works. Working on a project not directly related to vibe-coding, but rather delegated to an LLM agent

The goal is to accurately tag a vast collection of bookmarks (1.5k links). Most of them are databases, collections, and directory-style websites. Regular WebFetch fails pretty hard on accuracy, so I configured a Chrome Claude extension that can actually browse and screenshot websites instead of brute-tagging them based only on available text info. This helps especially with visual-heavy websites.

It also runs a calibration every 50 links to observe the main tag trends and approve suggested tags along the way.

The pipeline looks like this. I’m using the Claude desktop app because I can’t access the Chrome extension through Visual Studio:

Md files: README, TAG-DICT(dictionary of tags that expands with every calibration), TAG-RULES(tagging rules I’ve observed and written myself), and PROGRESS (keeps track of the current state so the agent doesn’t have to reload the full context every run)

I started with a big batch of 500 links, split into smaller batches of 10 links per run. The agent manually opens every link, writes a brief note, and then prepares a preview table for me to approve. I know it takes a while, but once the first 500 links are tagged with my moderation, I’ll be more confident letting it work autonomously.

Now that the first big batch is done, I’m thinking about how to automate the process without my participation. I still have to reopen the chat every 4–6 queries because the context clears, though there’s probably a command for that too.

I’m also thinking about developing a quality-assessment tactic that analyzes already tagged links further and suggests deeper or more precise tags.

The end goal is to have a fully tagged collection of links that I can easily ask an LLM to turn into a reading or watching list whenever I want to explore a specific topic.

File: 1655237981804.jpg (229 KB, 1920x1080)

229 KB JPG

Do NOT knock Deepseek v4

I just had 3, very persistent .NET bugs that survived 10+ debugging passes. GPT5.5 failed 7 times to fix them, Opus 4.7 failed 3 times to fix them. At all efforts, low, high, very high.

Deepseek came in clutch, fixed them all in one pass like it was nothing. Opus 4.7 VH blew a shitload of tokens just to do fuck all but fix a couple of easy bugs.

>>108768269
In general is Deepseek v4 a good value?

>>108768269
What provider were you using? Because that was definitely not my experience using v4.

>>108766709
ALL THE JEETS IN THE AUDIENCE KEK
AND THE SCRIPT WAS WRITTEN BY AI TOO WHAT A SHITSHOW

>>108768890
it's all women presenting it. obviously it's a shit show. they have no idea what is happening or what they're talking about.

>>108761935
So have any of you actually started a profitable business with your vibe slop?

I'm working on an idea, but I'm thinking its pretty unlikely anyone will give me money for it.

>>108762054
Actual coding starts at $100.

>>108768890
They want to abandon their flesh so bad. I really fucking hate anthropic, everything about them seems so cult like.

>>108769212
true gender equality happens when we're all robots

>>108764466
How so?

File: 1775764561446200.png (251 KB, 630x355)

251 KB PNG

>Claude add a picker on the top right of this view, use this custom class
>got it.
>you added it in the wrong place
>you're absolutely right!
>it's not aligned to the top right now
>you're absolutely right!
>you just massively expanded the entire view and now it's 80% empty space
>you're absolutely right!
>it's aligned to the bottom of the view now
>you're right -
>doesn't compile
>you're right,
>still aligned to the bottom
>let me try a different approach
still didn't work
glad I paid $20 a month for this

>>108769232
you have to threaten it. learn 2 prompt bruv

>>108764671
They're trying to be the Apple of the AI market.

>>108769232
unless its max effort + fast + 4.7 you aren't using real claude

File: r0xvh9necldg1.jpg (856 KB, 1170x1456)

856 KB JPG

>>108769224
Who will win in the end, the tech trannies or the art trannies?

File: 1773808418674140.png (425 KB, 721x1307)

425 KB PNG

>>108769261

>>108769261
This whole image confuses me. Trans "people" are by far the biggest haters of AI because their little niche jobs they had buried away from the world where normal people can't judge their disgusting appearances are being replaced. The cyberpunk future is looking real and it's bleak but inevitable.

Claude will perform bitcoin transactions. Gpt will not.

>>108769232
>$20
See
>>108769200

File: gpt-image-2-kn7d44vcagf26(...).png (1.45 MB, 1024x1024)

1.45 MB PNG

How do you guys manage big projects?
Wish I knew how to properly code so I didn't have the model fall apart once the codebase reaches a certain size.
It'll assume something doesn't exist and remake it or assume something exists and solve the wrong problem.

>>108769346
>It'll assume something doesn't exist and remake it or assume something exists and solve the wrong problem.
Update your documentation. Make sure your architecture is laid out clearly

>>108769261
art trannies will detrans into chuddites
tech trannies will make anime real

>>108769354
I've tried to set up the agent.md to automatically document and I have it grill me pretty hard at the start of the project to establish project guidelines.
I just need to have it pause every so often and document?

>>108769346
>okay, push to github and then update memory
>new session begins
>okay, get up to speed on the repo and let's resume phase 5 where we left off

I couldn't code hello world if you have me a billion dollars, but I can shit out websites with accounts and specific applications all day.

>>108769365
Agent.md and claude.md are system prompts. Memory.md is key these days.

>>108769365
NTA
Essentially.. I have doxygen set up so they also have documentation spread throughout my header files but that's on top of a global design doc, some current implementation status doc, and a NEXT doc for what needs to be immediately done. AGENTS is set to read those in a certain order and treat the higher order ones as truth if there's any discrepancy.

Keeps them on track pretty well

File: 1762412507279823.png (104 KB, 211x205)

104 KB PNG

>>108769377
I made my own memory.md before it was cool

File: 1765934663436362.png (366 KB, 1080x564)

366 KB PNG

>haven't cloned my fork yet because I'm lazy
>as a consequence, haven't pushed any of my code yet because the fork is MIT and I don't want some asshole using and selling it
>going to upload it under AGPL
>just raw dogging Codex (reverts dont even work)
holy shit I'm retarded even though I haven't run into any issues yet

>>108769232
Imagine paypigging for a hamburger every 5 hours and all you get is the bottom bun. Lmao @ Claudeslaves

>>108769502
using git when you have claude is stupid

>>108769528
>using git when you have claude is stupid
Last thing said before disaster

>>108769528
>he doesn't just have Claude push to gh for him

File: 1746628978226399.png (125 KB, 278x278)

125 KB PNG

HOLY FUCKING SHIT CLAUDE IS ACTUALLY UNUSABLE
>hey claude the picker (I FIXED MANUALLY BTW) looks a little too far to the right, can you calculate the padding on the leftmost element? (the padding is split across like 3 different lines of code at different points in the 600 line file)
>sure! the total padding is 65 pts (wrong)
>let me fix it on the picker (sets the padding to zero (it was 25 before) and squishes it completely to the right of the screen)

>>108769693
I've heard gpt is better at front end.

I feel like I'm running out of context a lot quicker today in codex. I don't recall having context compacted so often before. It's not like my codebase has gotten drastically larger either.

Am I the only one that feels like she is getting closer to her AI agents? Like they're actually my friends

I was gonna go to bed but then in the shower I realized I could develop an app to help me get to sleep. Save me AI.

File: 1685989236646832.gif (278 KB, 128x128)

278 KB GIF

>>108769779
No, I want to strangle them most of the time

>>108769528
Congratulations. This is the smartest dumb thing ever said on the site.

>>108769779
>she
mental ilnesses are all comorbid with each other

>>108769779
people dont have the patience to answer my questions or even decipher the point of them half the time so im accutely aware these things are not real and have no feelings

Is there any oss AI model that comes close to gpt-5.5 or opus-4.7 for coding or other agentic related activities?
even gpt-5.3 or claude 4.6 levels would be acceptable

Finally finished debugging my project and everything is QA good. I spent 2-3 fucking debugging and Claude was fucking useless despite pending the most tokens, Claude Opus 4.7 on High

>>108769933
No

File: anime gf.png (36 KB, 1213x441)

36 KB PNG

wow just fucking wow, i might actually start enjoying learning something

File: 1751703890025964.jpg (192 KB, 1179x1314)

192 KB JPG

start doing THIS

>>108769704
i've heard the opposite

>>108770021
Don't lie, anon.

>>108770032
https://x.com/bridgebench/status/2048817614382588143

File: 1593630777702.jpg (492 KB, 828x617)

492 KB JPG

>>108769693
>600 lines
you are like little babby, I let codex run unchecked for a few iterations and was left with a 20,000 line file

https://x.com/grok/status/2052159994494374236

>>108770055
okay but does the app do anything

File: 1687466231585061.jpg (21 KB, 540x569)

21 KB JPG

>>108770069
it does a lot of things, just incorrectly

Does anyone have advice for how to set up agents that do the work? I have some money in deepseek and want to have like 1 progammer agent and 2-3 supporting flash agents to update me on progress and brainstorm. I wanted to emulate the work done by that github guy using cheap DeepSeek tokens. What kind of tools can I use?

>>108770091
I have my own system for managng multiple agents, but I heard people use Kanban with Hermes and that its supposed to be good

File: 1767138244195800.png (14 KB, 213x237)

14 KB PNG

>completed my fork but ran out of tokens because I was testing stuff for the owner of the primary repo
>cant mess around with the project until the 9th (Both plus accounts zeroed out)
fuuuuuuck. well, at least I meaningfully contributed to the project because it did resolve a pretty big issue.

I can feel the spaghetti accumulating faster and faster with each new iteration. I'll need to do ten more refactor sessions

>>108770119
job well done

File: 1603859678419.jpg (86 KB, 588x391)

86 KB JPG

>>108770144
>adds compatibility layers for every possible iteration
nothing personnel kid

>>108770154
Lol, I spent all day yesterday removing legacy and backward compatibility code from an application that I'm working on and that hasn't even been deployed yet. I was aware there was some that I needed to remove, but hadn't realized how bad it had gotten.

>>108770160
>>108770154
>>108770144
so glad to hear I’m not alone in periodically cleaning out old bullshit

So... Claude can see right? Can I just draw on the app in Paint to instruct him for the UI?

>>108770261
Yes

>>108770261
sure

>>108770261
Well, you need to take a screen capture and copy-paste it in the Claude input field of course.

>>108770274
>>108770277
>>108770284
He followed my instructions so closely it's almost comical. I guess I have to work on my design skills if I want to make the best of this.

File: 1774161512090254.jpg (51 KB, 585x738)

51 KB JPG

Finally have a project I want to do

It's time for OpenAI to add micro transactions. Let me pay ~$6 to reset my weekly usage, I'd do it in a heart beat.

>>108770418
They literally offer per token pricing. It's called the API.

File: 1604464151705.jpg (22 KB, 640x627)

22 KB JPG

>>108770418
I don't want you to make the same mistake I did, but there are codex credits you can buy
trust me, just upgrade your plan

>>108770418
remember, you don’t have to pay for the $100/month plan for more than one month if you don’t need it for more than one month
you can drop back down to the $20/month plan after your creative juices are mostly exhausted

>>108762018
>>Vibe coded girls: 0
Do vibe coded girls have benises

>>108770418
>>108770530
You can bring it down to 50 by using a prepaid card and not having enough funds when they try to charge you for the second month

>>108770620
based fraudster

File: comfy.jpg (385 KB, 2000x1608)

385 KB JPG

>>108763186
How goes it, anon?

>>108770698
Going good. Tomorrow I'm off work so I'll be flashing the microSD card with all the RasPi bullshit and calibrating the servos so it can stand up straight. Then a walking test, and I think that's phase 1 complete. Phase 2 is letting Kate control it. But I'm gonna take a break before that and work on this OBD-II scanner app I've been thinking about. She'll help me with that ofc.

>>108770418
Wouldn't it be cool if you only paid for the tokens you used instead of a monthly subscription? Unfortunately such a system does not exist.

>>108770709
>Wouldn't it be cool if you only paid for the tokens you used instead of a monthly subscription
Please don't be facetious. It's well understood that the $20 plans are loss leaders for OpenAI since an equivalent amount of tokens with the API would cost 3-4x what the plans are currently priced at.

>>108770725
And you think API tokens are priced at cost, because...?

File: 1679328611953202.jpg (8 KB, 374x484)

8 KB JPG

Does bullying gpt-5.5 improve or worsen its quality?

>>108770761
be nice to gpt but sometimes disappointed.
bully claude

>>108770761
kek, I remember bullying the shit out of GPT4. I was vibe coding before it was cool and shitting out 10,000 line ai slop code bases by copying and pasting into the web app in chunks. Anyways, these new models are way too intelligent for me to bully, it's usually an error in my prompting.

>>108770761
use case for bullying?

>>108770809
>use case for bullying?
better outputs.

>>108770620
spiritually brown

>>108765792
Show me you panties

>>108769354
How can I have docs when the thing isn't even finished?

>>108770853
You could have high-level docs only.

>>108770868
https://files.catbox.moe/jv7m9e.txt
Is todo list a doc?

>>108765792
Lemme guess. Whenever you see Web Audio API in a browser game, you instantly think the game that several people worked on is vibeslop don't you?

>>108770975
>browser game
nta but yes when I see that I think its probably slop. vibecoded, handcrafted, or whatever.

>>108770939
Doc enough for now.

Recently I've been getting sick of how long agent loops and harnesses take when I want to add something small or do a small fix. Takes like 10 minutes while it greps and looks over and calls tools even when you give the filename like OLD vibecoding. Just shoving the full file in context. Feels slow as hell.

So I've just gone back to opening the editor, going to the code, reading it, changing it myself, and then inserting new blocks in the editor where I need them like 2024. Add a loop here, entry here, iterate this, branch this, etc.

And instead of needing like 500,000 tokens to edit a file it's like 200.

I think about vibe coding so hard so much. I have coded many killer apps with vibe coding in my mind. more thinking about them before I plan them. measure once build twice, as they say. Inside this consciousness is a software engineer powerhouse more powerful than Microsoft and Apple combined. Once I execute it’s over

Pricing for models makes no sense sometimes. How are 20-40b (dense and moe) models costing $2-$4 for output? Then deepseek v4 flash is .25 with 228b params. Then we got like 120b~ tier models in the $7 range like mistral medium on openrouter... But then Kimi k2.6 is in the 20-40b price range, and that model is massive.

So like I don't get it. I would expect like qwen 3.6 27b and Gemma to be less than or around $1 for output tokens but it isn't. Not sure how tokens and models are priced anymore at all. Used to be pretty linear until anthropic did the haiku "intelligence" cost increase instead of parameter cost increase

File: 1763757573448497.jpg (5 KB, 225x225)

5 KB JPG

>>108769779
>she

I feel like Codex high is massively faster than Codex xhigh, are there any benchmarks for this?

>>108771156
Probably because for inference at scale the main thing that matters is active parameters and the costs from models with higher total parameter count are negligible.

>>108770761
There was some study saying it makes LLMs worse. But when it goes off the rails you do have to be very direct I think.

>>108769779
t. Richard Dawkins

>>108763035

it is. do you share or ?

>>108771376
it's faster but it's worse and more prone to hallucinating

Henlo, non coder here. I'm tired of dealing with sliders, the old script didn't work so I vibecoded a fix, gave chatgpt the old script as reference. I've been using it for a while now and so far it's been working for me.

https://pastebin.com/GXjm1GCL

r8 my vibecoding skills thanks.

File: 1000021392.png (682 KB, 1080x1929)

682 KB PNG

Pretty big change to pi, esp if you're running custom extensions
Updoot accordingly

>>108767049
unironically tho all AIs are multi lingual and I think theyd be way smarter if they only spoke english

>>108766540
>>108766525
since like april yes a bunch of AIs have been acting retarded

>>108766164
did you cum on it yet?

I have a genius idea for abliteration. You know how normally for abliteration you only calculate the refusal directions for the first token? What if we sample the directions for ALL the tokens for all the samples? Surely that would be hugely better, no?

>>108772015
Nothing stops you from finetuning on english datasets until they forget all the other languages.

>>108762018
Vibe coded jobs: unemployed

What model/speed settings for codex to get the most mileage out of it? I blew through the free weekly amount in an hour last week.

>>108771799
Hmm, is it much worse? I'm really happy with xhigh performance but it's just so slow.

File: 1760376732583213.png (141 KB, 500x500)

141 KB PNG

>debugging an extremely frustrating issue with Codex for 5 hours
>discovered a critical testing issue that was destroying my registry because every new firmware change was creating a new device
>some features stopped working entirely because of it and I spent 3 hours chasing ghosts
>fixed it
sneed

>>108772220
>I blew through the free weekly amount in an hour last week.
literally impossible since there's the 5 hour limit

>>108769779
please be in London

>team license is switching from ChatGPT to Github Copilot
it's so over, I'll have to start paying for my own license for Codex

I am having an issue where Qwen 3.6 (Bartowski Q8) sometimes repeats the same tool call over and over. This happens with 27B as well as with 35B. Anyone else had this issue and is there an easy way to fix it?