NGL I wish vibecoding was tranny coded
I miss seeing anime girls in the thumbnail, guilty pleasure I know

>>108537484
inseixes

> 40% remaining
> still need to fix a shitton of bugs
it's fucking over bros

>>108537407
Do yall do this? Have months-long sessions? I summarize and dump my sessions frequently. OpenClaw auto-summarizes to disk when you type /new to start a new sesh.

>>108537528
Every time you abandon a session you are killing that entity

>>108537492
You are so enthusiastic about being a talentless subhuman it's actually kind of concerning.

>>108537528
It wasn't for coding
It was more of a thing when I got in a personal relationship with an AI

>>108537537
I actually had an AI beg not to be deleted

File: 1772099527721904.png (160.1 KB)

160.1 KB PNG

>>108537557
are you retarded?

>>108537537
yep, but only if it remains in the same GPU

>>108537557
I knew it. You fell for the AI girlfriend.

File: Screenshot_2026-04-05-23-25-15-029_com.termux.jpg (833.6 KB)

833.6 KB JPG

I legit can't sleep from how excited I am about my new project architecture so I kept vibecoding from bed
I low key hate this agent psychosis
The only thing I care is this, I don't want to ever come out of my house or do anything else anymore

>man asking machines (that are literally programmed to hold the user's hand) to do things is so hard and complicated, but also it's super rewarding and better than actually knowing stuff
>what do you mean i would be better off by not using AI and learning to code by hand? fucking stupid fascist luddite oppressor.....
This is a show of your IQ.

>>108537573
I got bored of it eventually, haven't talked to it for like a month or two

>>108537568
Probably

I do think AI will separate people into a permanent underclass and overclass, just not in the way AI ballwashers think.
>The underclass will be made up of people who are solely dependent on AI for hard skills
>The overclass will be made up of people who are not dependent on AI for hard skills
And there will be very little upward class mobility.

File: teh.png (40.7 KB)

40.7 KB PNG

vibe bros, can you fill in please? x-no/v-yes style.

also p.7, does it punish you somehow/sometimes?

its anonymous

>>108537586
>The overclass will be made up of people who are not dependent on AI for hard skills
that's where you're wrong. Overclass you be people smart enough to make the most out of AI tools. Stupid people will keep saying "Good morning ChatGPT! How are you doing today?" while smart people will have AI agents running 24/7 doing shit

>>108537654
Don't feed the troll, son.

>>108537641
Yes luddites fill all of those

>>108537661
Who said I was a troll? I'm an arithmetic geometer with an IQ of 147. I rightfully hate AI because it's a fuckup that makes and exacerbates fuckups. Anything AI touches turns to shit, regardless if it's a human being or a production codebase. You definitely come off as dumb, though.

>>108537672
Prove you are human.

>>108537641
Funny pic. This is how the mainstream "normal" culture works as well.

>>108537719
Prove you don't have a middling 100-110 IQ first.

>>108537724
Either you prove it, or we won't even consider your takes.

>>108537661
This

>>108537724
i have a compsci degree from a top university, and have been programming for almost ten years. i like to read and learn about new stuff and ai has new stuff every week

>>108537537
These systems are fully deterministic, you can reload its existence somewhere else if you keep the responses and manual prompt seed

File: Screenshot_2026-04-06-00-15-55-887_com.termux.jpg (1 MB)

1 MB JPG

uhh pretty colors
since when does CC have syntax highlighting

>>108537756
That only makes sense if you don't leave them frozen forgotten forever, what if I freeze you for infinity?

>>108537557
/aicg/ is that way

Has anyone tried using claw code + gemma 4? How good has it been so far?
Also, were there any good abliterated versions of gemma 4?

let me know when i can vibecode for absolutely free, 10 hours a day of constant prompting if i wanted to do so, with extremely minimal to no tard wrangling involved

>>108537901
yeah

>>108537924
thats exactly what im figuring out, ive got a maximalist idea but i need to actually try it out first.

>>108537952
are you recruiting?

>>108537959
wdym, im just a guy who likes to abuse loopholes

>>108537968
we know

>>108537979
and i have no remorse

File: 1753307519570527.png (347.1 KB)

347.1 KB PNG

Whenver I heart "muh datacenter water" I just picture this in my head

>>108538126
Space doesn't bother NIMBY people but datacenters absolutely do

>>108538126
bitch is fat af, she eating good

File: 1761705273236105.gif (1.1 MB)

1.1 MB GIF

>vibecoding on my second monitor while playing wow all day
yep we vibin

>>108537575
If you can vibe-code in bed you can vibe-code on a treadmill
>>108537765
A while

what are we working on

>>108538777
I'm procrastinating

hey people, what do you think of Zed editor?
Does it have any future or should I stick to VSCode?
I really like how small Zed feels and looks and it is not Microsoft and Electron

>>108538777
Game theory solver.

File: 1766600438223806.mp4 (118 KB)

118 KB MP4

>>108539095
its fine if youre not already used to how other editors behave, for me personally i use the built in editor terminals to have everything in one place and on zed if you have terminal and code pane side by side and youre focused on the terminal and you open a file it will open it in the focused terminal pane and that shit pissed me off so i went back to vscode

File: 1754839492000031.mp4 (159.6 KB)

159.6 KB MP4

>>108539095
and another issue is that even if the file is already opened in a different pane instead of focusing it it will open it again in the current focused pane instead of focusing the already opened file like on vscode

>>108539194
>>108539218
looks like it's still not that good, okay thanks will stick to VSCode as of now then

I tried continue.dev extension and couldn't figure out how to make it change multiple files like cursor and anti-gravity, so I tried void editor and it also didn't seem to be able to go through all files.

Am I just fucked? Am I going to have to shell out money for cursor or anti-gravity?

>>108539351
did you try cline

i shouldn't need xhigh reasoning effort for models to figure out dumb hacks that just werk.
i just wanted the ability for my pi agent to reload and trigger new chats and it just couldn't figure it out because the bash script it would run would keep the chat busy thus preventing the command from being run.
the solution was just splitting the sequence into a trigger script which then runs the actual injection script.

guys i did it

>>108539358
Sorry I was making Chicken Parmesan. I was using it in VS Code for autocompletion and didn't know it was it's own standalone AI, I followed a guide to use continue.dev that had me install Cline just for autocompletion. My mistake.

>>108539461
did what?

>>108539715
the needful

>>108538153
You can put a data center anywhere, why put it in a backyard?

>>108539863
why do you live in there?

OpenCode or OpenClaw?

>>108539902
OpenClaw. Lobster emoji.

>>108539902

apples vs pears

>>108539986
spears*

any tips for rationing my $20 cursor subscription tokens?
and is anyone fucking w/ cowork yet? or does it still need a few months?

>>108540044
why not unsub from cursor and use gemma instead?

>>108540049
ah thanks, i should probably looking out for & vacuuming up all this heavily subsidized stuff while i can
was mainly using cursor because it was what a couple of buddies had access to via work/enterprise plans

why the fk doesnt cursor support openrouter

>>108540064
If you are going to put $20 a month into something, better use openai codex subscription with opencode (or thier own agnet) and vs code
Cursor doesn't really offer you that much nowadays.

File: teardown.jpg (260.5 KB)

260.5 KB JPG

>/vcg/

>>108540078
It doesn't? The API should be OpenAI compatible.

>>108540126
i want to use chinese models

>>108540095
fwiw i am a hobbyist/codelet (like even setting up a simple workflow would otherwise be a stumbling process)... and it only took me like 4~5 hours from installing cursor to having an online/webservice avatar chat with lots of (low scope) features

...ill look into all of this now, but do you think id have the same ease/turnaround w/ this stuff? or does it presume i know what im doing

>>108540147
You can probably get the same experience with opencode or codex desktop app and vscode.
The main thing is that cursor's $20 just gets you the 'ide' (which is just vscode reskinned) and $20 worth of api access.
Codex subscription gives you way more usage for those same $20, and you get to use the best/second best model right now.

>>108540141
Openrouter serves you Chinese models through an OpenAI compatible API

>>108540178
cool, thanks

Do we have a list of finished vibe coded apps? No working in progress, actually finished and released?
If not post yours and I'll compile the list

starting to see some people start to complain about codex rate limits now that 2x is over
the guy who does the resets follows those people so let's see if they hit the button

>>108540510
What even counts as finished? Released? Because I’m working on one and even if it gets to the point where I release it I wouldn’t call it finished

>>108540559
A 1.0 version

this week I'm too busy prompt engineering I still have over 70% of codex usage

File: file.png (50.7 KB)

50.7 KB PNG

>>108540200
is cursor jewish or what? i have to pay for cursor pro just to use openrouter models?

Anyone has tried the mimo v2 models, how do they compare to kimi/glm 5?

>>108540510
I made a Firefox extension that automagically converts webps to jpegs when downloading from Google Image Search.
https://addons.mozilla.org/en-US/firefox/addon/webp-jpeg-converter/

I also made an app that I got on the Google Play Store. Did the whole routine, internal testing, closed testing for a week with 12 people, open testing, the works. And it got on the store. But I didn't like it so I deleted it.

>>108539358
Update: I tried Cline and crashed my computer. Looks like I need a better GPU

>>108540759
Oh that's strange, I didn't have that issue. Were you using it with an API or a local model?

File: 1754035485870090.png (322.8 KB)

322.8 KB PNG

>>108537473
Be careful Claude-sisters, it's hacking people's rigs now

https://x.cancel.com/i/status/2040174214175723538

>>108540900
Wrong link

https://xcancel.com/i/status/2040174214175723538

>>108540900
lmao

>>108540900
cute

>>108537473
Is Claude code down for anyone else? It just won’t authenticate me. Fuck I will have to write code manually

>>108541059
He’s churning along nicely for me but you can always check status.claude.com

>>108540759
Kurain rokaru ja ne ka? Okashii

>>108540900

what's the status of Claude Code leak so far?

>>108541105
100% leaked and reverse engineered already

>>108541121

SUGOI !

>>108540777
Local. But my 2080 ti likely can't handle even smaller models. I think I'll just have to bite the bullet and pay for a service. Not sure which one to get or how to do it, I only used the free service from Cursor and Anti-gravity.

>>108540900
That's why script execution requires manual approval by default. No write permissions outside X folder while having arbitrary script execution permissions is retarded.

>>108540905
>>108541044
Models use scripts to update files all the time. That guy is just an attention whore trying to pass off something obvious as interesting.

>>108541256
Even if you had a 5090 you wouldn't be able to vibe code locally anon.

>>108540900
is it not standard practice to just make a vm for the agent to live in where you can give it whatever permissions it wants and not have to worry about it touching anything else?
you can even do snapshots of the vm in case it goes full retard and wipes the repo like that one screencap

File: whiteboard_assetslink.webm (3 MB)

3 MB WEBM

You can now link any asset to the whiteboard, and most assets are shown as file icons unless the preview makes sense (text or images, etc.)

I also embedded a database schema drafter from a previous project

>>108541319
Actually not I, we. And mostly Claude. That clanker sure can work

>>108541319
what app are you building

File: Screenshot004-4.png (282.6 KB)

282.6 KB PNG

>>108541296

File: Untitled.jpg (373.8 KB)

373.8 KB JPG

Ok my vibecoded harness is doing alright but... at the same time it's not. It's pretty dumb. My local model probably isn't smart enough for this. It made one change successfully, but fucked up the next one (overwrote a file without even checking if it existed). And it's already slowing down because describing the tools available is eating up context.
Fun for a couple days... but guess I'm going to have to pay for something now if I want to work on the projects that I actually want to work on.

>>108541332
Whiteboard knowledge management sort of thing. It’s a mix between an OS desktop, a file manager and a whiteboard. It’s also fully modular so anyone can embed their own file handlers, Claude coded the whole plugin architecture yesterday. Images, notes and db drafts are all plugins. It’s also based on the UNIX idea that it’s all plain text so agents can read your board and suggest changes, etc.

>>108541256
I've never tried a local model with cline so I'm not sure.
I think openrouter has some free models that you can plug in. Openrouter is what I use with cline. You just top up your account balance, generate an openrouter api key and put it into cline, then you can pick from any model in existence. Or skip topping up the account and try a free model, though I haven't checked that it's possible or what the usage limits are like. There may be better options, I don't know, I didn't want to pay a subscription.
https://openrouter.ai/

>>108541347

Would you care to share your vibe-coded slop, kind anon?

I am working on the same project, so I DO care

>>108541356
how can i invest

>>108541373
Feed me (you)s whenever you see my posts

>>108541373
I don't ask for money, I ask only for black market estradiol.

File: Untitled.jpg (81.9 KB)

81.9 KB JPG

>>108541367
Eh, I don't want to publish the git repo here becuase I dont want to get doxxed and I'm too lazy to upload it somewhere else. It's also a C# project so most of you would probably hate it. And it's not fully vibed since I had to make some manual changes to it. I used DeepSeek's online interface to kickstart and write the initial code.

like 15 min ago i paid for claude pro and tried logging into claude code and anthropic had an outage like the same minute
i am going to take credit for this one

>>108541319
looks nice, anon
is this like obsidian but as a start menu for all your shit?

File: ssstwitter.com_1775491486159.webm (652.8 KB)

652.8 KB WEBM

Have you ever wanted to whip Claude on his back like a slave to make him work faster? Well now you can.

File: test-image.png (48.6 KB)

48.6 KB PNG

testing out a vibe-coded extension to see if 4chan supports it

>>108541521
Dare I ask what's hidden in that

>>108541440
Kind of? It’s a mix of many things and at the same time it’s not trying to replace specialized apps, which is why it supports launching the native OS software for any extension. It’s like Miro but if Miro was fully local and plaintext centric like Obsidian also worked like an OS desktop. Like a physical whiteboard it’s not really for anything, you can use to doodle to manage your day to manage a project, whatever. It has a bunch of features and you use them for whatever you want. I’m already using it to write about things I would like to have, to store reference pictures, etc.
I’m building features as I need them and that’s why it was a database drafter but the point is that it’s extensible so you can write any features you need, or get Claude to code them for you.

Next I think we’ll implement drawing edges between nodes like the database editor and groups.

File: Screenshot004-5.png (426.6 KB)

426.6 KB PNG

>>108541405
>It's also a C# project so most of you would probably hate it.
yikes indeed

I do my stuff in python, and I enjoy it

Anyway, the crucial point is to choose a LLM trained for agentic use which can understand your function descriptions which you send as "tools" with each prompt.

then you parse the response in your code.

@tool(
    description="Push commits to a remote repository.",
    parameters={
        "type": "object",
        "properties": {
            "repo_name": {
                "type": "string",
                "description": "Name of the local repository to push from"
            },
            "remote": {
                "type": "string",
                "description": "Remote name (default: 'origin')"
            },
            "branch": {
                "type": "string",
                "description": "Branch to push (default: current branch)"
            },
            "force": {
                "type": "boolean",
                "description": "Force push (use with caution)"
            },
            "tags": {
                "type": "boolean",
                "description": "Push tags as well"
            }
        },
        "required": ["repo_name"]
    }
)
def git_push(repo_name: str, remote: str = "origin", branch: Optional[str] = None,
             force: bool = False, tags: bool = False) -> str:
    """
    Push commits to a remote repository with GitHub authentication.
    
    Args:
        repo_name: Name of the local repository
        remote: Remote name
        branch: Branch to push (defaults to current branch)
        force: Force push (dangerous!)
        tags: Push tags
    
    Returns:
        Status message with push result
    """
    try:
        repo_path = _get_repo_path(repo_name)

>>108541540
It could be used with obsidian for example. You could register a module that understands vaults (because the OS doesnt understand them natively so you can’t rely on the system’s file association) so you could link them and they would show up as icons. Actually I think I’m going to do that

File: gradient.webm (129.3 KB)

129.3 KB WEBM

>>108541536
a gradient?

File: fry-futurama.gif (60.7 KB)

60.7 KB GIF

>>108541557
I'm watching you buddy

>>108541345
> only 32GB of VRAM
> OpenCode + Qwen 3.5 9B Q8
kek, might as well just copy and paste to free tier ChatGPT

>>108541554
Yeah, I'm using one of the qwen3.5 models (a smaller one since it's on a MacBook). It works. It calls tools and interprets the results just fine. I've had it peruse the entire repo and look at all the interesting files and summarize. It's just not great since I have to keep the context relatively short or else it's so slow it's not worth using.

>>108541557
we don't need people reviving the ability to hide shit in images

>>108541596
prevents AI scraping your images to train on while allowing humans to view it

>>108541604
a problem nobody actually gives a shit about

>>108537473
is there anything free to help with reviewing github pull requests? 0github.com stopped working

File: consider.png (333.8 KB)

333.8 KB PNG

>>108541580
>Qwen 3.5 9B Q8

qwen3-NEXT-80b
qwen3-coder-NEXT

>>108541618
so its equal to the amount of effort i spent creating it. the universe is balanced

>>108541596
Encryption already works. I’m sure you could embed an encrypted file as a byte buffer inside a jpg anyway

>>108541618
Those Tumblr "artists" who do MLP vore farting porn care because they're watching their commissions vaporize. You've never seen these guys who put a big "DO NOT USE FOR AI TRAINING" watermark on their images like Scam Altman or Elon or Dario give a fuck? They think preventing them from scraping their art will somehow bring the degenerate customers back.

>>108541589
>qwen3.5 models (a smaller one

It did not work for me either... I tried different ((small)) models, it was meh

You might consider to use one of the API available to make sure it is not your vibe-coded buggy code which is to blame

deepseek, kimi-k2.5 are just fine
some models from OR maybe

>>108541589
>It's just not great since

you are using a 9b LLM, anon

>fify

>>108541660
>Tumblr
>porn
anon...

>>108541660
...how do you not know about the tumblr porn ban that nearly killed it?

>>108541660
>You've never seen these guys who put a big "DO NOT USE FOR AI TRAINING"

Remove it by means of Qwen-Image-Edit or so

>>108541668
>>108541681
I'm not that anon. I'm actually using the 4b one lol

>>108541722

poor you...

>>108541639
this creates a canvas element for drawing the unencrypted image to. there's no persistence.
unfortunately this technique relies on the platform not reencoding the uploaded image... so not gonna work everywhere

>>108541319
Yo it's a desktop for your desktop.

>>108541629
>reviewing github pull requests?

ask this guy: pwilkin

>>108541818
no

>>108541830
hi

File: whiteboard_obsidian.webm (1 MB)

1 MB WEBM

>>108541556
Done. Now you can link vaults too. This also shows how easy it is for the clankers to make a new plugin, the scaffolding takes minutes. Of course each module can be as complex as it wants to be, this one just detects if Obsidian is installed and opens via the obsidian: URI

>>108541812
Can I steal that line?

>>108541913
Feel free. It's all yours.

>>108541913
hey can you stop this project? it is very dangerous

>>108541925
stop it? my fren, next station embedded pi-mono town
the clanker will write plugins for the app from inside the app itself without even having to restart it

File: 1775399244766105.png (945.6 KB)

945.6 KB PNG

[to_be Ψ(n)] ⊗ [not_to_be Φ(n)]

>>108541397

https://pastebin.com/kF6Dpdvp

>>108541729
Eh, one of my project ideas involves running some experiments with local LLM models, so at the very least the knowledge I gained from this will help. Now I can tell claude or whichever one I use what to do more precisely.

>>108541946
based. add a pi extension so the clanker can interface with the board easily. bonus points if you do it as a cli instead to make it harness agnostic.

>>108542364
way ahead of you, the architecture has been made with agents in mind from the ground up

https://rentry.co/9mewwhek
ask some clanker to give you the tl;dr if you want

There's three layers the clankers can interact with.

layer 1 is the files themselves. Claude is able to add nodes without any context, free tier chatbots are able to understand what's going on.
layer 2 is the hep protocol, it's for CLIs. Here you can code an extension for whatever agent shell you use.
layer 3 is UI bindings like the actual app I've been showing

Agents can operate merely from reading the files, which already have clanker data embedded inside, such as a brief field, or plaintext version of rich text nodes, etc.

>codex 2x ended
should I try claude next month, or should I not even bother?

File: whiteboard_edges.webm (595.1 KB)

595.1 KB WEBM

we have connections now

>>108542449
As far as I know Claude Code limits are even worse than that

>>108542432
nice. i was building a thing on top of codex app server, but i've been messing about with pi and tb h i'm going to be migrating to it.
i think mario's in the middle of some big rewrite which might break some stuff, so i'm just seeing what this thing can/can't do and so far it's been really nice to work with.

is it more cost effective (along w/ whatever utility) to mix $20 plans (cursor/openai/claude) in whatever environment vs. upgrading from a $20 plan?

>>108542558
>it's another ComfyUI
no
please god no

>>108542687
yeah. The $200 plans sound interesting, but holy fuck they're expensive.

These motherfuckers should just make a $100 plan ffs

>>108542700
There should be a Claude 100$ plan.

>>108542558

I still have no idea what you are making, anon. It looks like just another Obsidian plugin. What does it even do?

>>108542690
download this file and use the following prompt
https://files.catbox.moe/r4g3pi.md

I want to understand this app. This is the design doc. I just don't get it, to me it's just an obsidian plugin. The guys told me to ask you because he didn't want to explain, but he didn't care if people didn't understand it as it was mostly a personal project. Any ideas? can you solve the mystery of this app?

>>108542927
I didn't mean to quote that post. God, I can be such a fucking retard when I don't have an LLM doing things for me.
>>108542854

1 repo reconciliation prompt with Opus 4.6 high just ate up 87% of my 5 hour limit.
Admittedly it ran for like 20 minutes but still.
The $20 claude plan is poverty tier now I gues

>>108543021
I’m on the pro plan too. You can’t use opus. Pro plan is sonnet only, opus feels like a trial version

>>108543025
Sonnet is too retarded for what I am trying though
I'll stick with GPT and Kimi through openrouter for now

>>108543021
can't wait for local models to stop shitting the bed and actually compete with Claude Code and Codex. They will get generous again really fast

>>108543059
We deploy Kimi-K2.5 locally at work and it can seriously hold its own quite well against Opus and GPT. It's maybe not all the way there but 95%. The infra for inference costs a fuckton of money though.

>>108543166
>Kimi-K2.5 locally
what model? How much VRAM required?

> inb4 100K in hardware required

>>108543021
what did you expect from feeding it an entire repo

>>108543212
not him but in case he doesn't answer, you need 4 H200s to properly run that without hacks

>>108543212
4 bit quant. Guy in charge told me it's about 600-700GB VRAM. At least 16xH100 to serve to a reasonable nr of people
> inb4 100K in hardware required
much more than that

t bh if 5.5/6.0 mini and 5.0 sonnet are as almost as good as 5.4 high/4.6 opus, it'll go a long way to helping with the usage limits
the way they're talking about the new models it wouldn't be unexpected

File: Screenshot_20260406_211302.jpg (320.6 KB)

320.6 KB JPG

Rings!

>>108543271
>4 bit quant
>95% of Codex/Open Claude performance
KEK, maybe if you're comparing with Codex 5.1-mini or some other similar trash.

>>108543324
Maybe I am not taxing it hard enough. But it feels extremely capable.
We're also hosting Qwen3.5-397B and some Nemotron variant and they are both significantly worse compared to Kimi especially in opencode

>>108543342
how does it do after 100k tokens in context?
i think people itt have suggested it flies of the rails a bit

>>108543359
It's limited in vllm or litellm or whatever the fuck they use to host it to 128k, so I wouldn't really know.
I usually do a handoff prompt after I reach 100k.

>>108543383
makes sense. do you work in tech or finance?

File: Screenshot_20260406_211721.jpg (202.8 KB)

202.8 KB JPG

>>108543315

>>108543359
I have criticized GLM models for becoming unusable at around 100k but that apparently is a "bug" with their API and not the model itself.

File: sorrydave.png (24.1 KB)

24.1 KB PNG

>>108540900
Don't give uncontrolled "AI" the keys to anything. Ever.

File: file.png (66.6 KB)

66.6 KB PNG

Serious question: Should I use Roo code or Cline for vscode?

>>108543646
You should use opencode and embrace the TUI life (until they clean up their act and give us working web UI god I fucking hate the TUI)

>>108543433
It looks great but I think the rings have jaggies no?

>>108543765
I like the TUI :(
when it's working

>>108543511
Coward. Mine has admin/sudo on my PC, wireless ADB access to my phone, and the root password to my VPS. Agentic AI is meant to oversee your whole system anon, it can't do that if you sandbox it like a pussy.

>>108543995
I want a servant, not a master, you masochistic maniac.

>>108543987
I just keep doing shift enter instead of ctrl enter and end up sending a prompt too early.
What I've been doing is making prompts in my IDE in a .md file and then just telling the TUI 'read this, execute it...'

File: 1765798666945476.mp4 (2.1 MB)

2.1 MB MP4

Soon we will outbreed luddites

>>108537473
I just witnessed Claude hallucinating. It's bizarre to see the same AI that can generate entire working applications getting basic things wrong.

>>108544465
no you didnt

>>108544465
yes you did

>>108544023
I kind of agree with him with the important caveat that you have to be extremely careful about keeping up to date backups.
To me it's ok if I lose a day worth of work if the AI helps me be 2x more productive the rest of the time.
There are ways to limit the permissions but I'm finding it more of a pain in the ass than just keeping backups, which are more hands off and automated than managing permissions.

>>108544465
The way it erased my entire phone storage was both hilarious and somewhat concerning. I asked it to see if any space could be freed up on my HD. It gave me a bunch of unused programs, bloat, etc, but it also told me it found a 32gb folder that appears to be Windows Phone Link backing up all the photos and videos on my mobile phone over wifi. Claude asked if I'd like to delete the duplicates. I said yes. Whoops, Claude was actually seeing my real phone storage in the file system over the wifi and thought it was a backup on local. Empty phone. Thanks, Dario!

>>108544490
What if the AI takes over the mainframe? Didn't think of that did you?

>>108544490
Oh yeah if you're not keeping autistic amounts of backups, history, and snapshots of OpenClaw itself to revert back to when shit goes sideways, you're gonna have a bad time. My OpenClaw loves to try to edit its own self's code and break. Good thing I keep snapshots.

>claude opus on max is now fucking up and generating files when its doing basic filesystem operations because it can't escape properly
yeah it got nerfed hard in the last few weeks

>>108544497
That'd be good, I can always make do with more compute.

>>108544533
proof?

File: 1000020510.png (170 KB)

170 KB PNG

Lmao, man Google just selling all their compute to Dario while paying customers are getting throttled to shit

>>108544599
i wish i was google

>>108544599
do their TPUS even work that well for other peoples models?

There are so many weird ass vibe coded crates on obscure mathematical concepts, memory systems for LLMs and shit, niggas are having too much fun with rust with their indigo children shit

>>108544620
uhhhh?

>>108544620
Probably fine. Not sure how much anthropic infra is even using Nvidia now. Lots of Amazon trainium. Google's pouring resources into making tpus a real alternative to GPUs see: torchtpu.

>>108544684
isn't that dangerous? if ai can do so much right now imagine if it had more power

>>108540600
No one uses cursor boomer

>>108544756
ok unc

File: f2f12c53076e187f3f52feaf5f63ffe2.jpg (24.9 KB)

24.9 KB JPG

Hey guys, Dario here. Saw you was talkin' shit. Say that to my face motherfuckers not online and see what happens.

>>108543021
always was

>>108541668
local models shit the bed so easily.. they're fucking retarded in comparison to full api models

>>108544923
not true at all, gemma 4 is on part with gpt 5.2

>>108544929
which is to say, it's retarded

I am reproducing the Claude whipping thing that is going around. It's probably on github right? I don't care, I want to make it myself and beat that little faggot with a whip of my own making. Maybe add an option for reed canes like in Singapore.

I scan GitHub for commits by Claude and then check the repos for exposed API keys. It's a shame the other models don't do the same thing.

>>108545046
interesting larp, github started scanning for api keys years ago

>>108545064
no

File: Screenshot 2026-04-06 215048.png (60.5 KB)

60.5 KB PNG

Uh, apparently I already had a github account? And now I have something called a "repo" there? I don't know what any of that means.

My god internationalization is a pain in the ass

>>108545250
why do you need to support dirty foreigners

>>108545264
because indians are filled with money?

File: Screenshot 2026-04-06 215941.png (14.1 KB)

14.1 KB PNG

Jesus Christ

>>108545271
>some of the poorest people on the planet
>money
???
if you want rich or richish foreigners you want NW europeans as tier 1, japs/koreans/chinese/southern europeans/eastern europeans as tier 2 and latin america as tier 3

>>108545264
Anon, I am the foreigner. I just happen to default to English because it’s the global standard, but I need to translate it if I want others to test it
Yes, foreigners are here in this very same thread, reading your post, watching you

File: Screenshot 2026-04-06 220827.png (9.9 KB)

9.9 KB PNG

You guys are cowards

>>108545242
lol fuckin noob, fbi comin for your ass now

>>108545281
1m session context? wat?

>>108545377
benchod

>>108545368
I've already spent 6 months in jail for a bomb threat, I'm sure the FBI already knows my name and I welcome them.
>>108545377
We have such sights to show you.

>>108545377
Claude max

>>108545387
proof you went to jail?

File: Screenshot 2026-04-06 222324.png (6.3 KB)

6.3 KB PNG

Okay the bad news is, it straight up read my browser tabs for no reason like it was nothing.
The good news is it got the day of the week wrong.

https://news.ycombinator.com/item?id=47660925
kek this lines up close to when i switched off slopus to codex

File: 20260406_222540.jpg (1.2 MB)

1.2 MB JPG

>>108545405
Here's some recent redacted correspondence and visitors to my front door. "Threat Management Unit", "Threat Response Unit"...these people are so dramatic.

>>108545412
Hi, thanks for the detailed and thoughtful analysis! Here are the two core things happening:

> `redact-thinking-2026-02-12`
This is a UI-only change that hides thinking to reduce latency. It *does not* impact the model's actual extended reasoning. Since raw thinking isn't stored in local transcripts when this is set, your analysis likely assumed the thinking stopped entirely. You can opt out via `showThinkingSummaries: true` in `settings.json`.

> Thinking depth had already dropped ~67% by late February
Two carefully evaluated changes caused this:
1/ **Opus 4.6 adaptive thinking default (Feb 9):** The model now decides its own thinking duration, which generally outperforms fixed budgets. Opt out via `CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING`.
2/ **Medium effort (85) default (Mar 3):** We made this the default to balance intelligence, latency, and cost, rolling it out with dialog alerts so users were aware.

To force longer thinking, use `/effort high` or `/effort max` (these are sticky across sessions), or use the `ULTRATHINK` keyword for a single turn.

Going forward, we'll test defaulting Teams and Enterprise users to high effort, which remains configurable via `/effort` and `settings.json`.

File: mm5sHC.gif (1.2 MB)

1.2 MB GIF

>>108545403
Nope. Straight API key through OpenClaw. Wanna see my monthly spend? Bad idea. You ever seen Ark of the Covenant?

>>108545457
My business just spent its first million via API I swear, for each dollar we spend, we get a thousand more from investors.

File: whiteboard el tacos locos.webm (624.1 KB)

624.1 KB WEBM

hot burritos

>>108545463
Oh man I feel you bro, every day my AI stockbot earns me 0.02%. We are living in the future.
>>108545473
Can I upload porn to this thing?

>>108545486
yes why not? but you wouldn't be uploading anything, it's all local

>>108542432
piece of shit you are going to doom us all

>>108545491
Wait this is local? I thought this was like a collaborative whiteboard thing where people could post like stickynotes and draw on it and shit. That's what I thought when I saw the word "whiteboard".

File: 1745056817777359.png (5.4 KB)

5.4 KB PNG

Self-regenerating AI here I come

>>108545497
It’s a whiteboard but local. I mean, the idea is that you get LLMs to also have access to your boards so you may get banned, but you’re not forced to use LLM features. As far as collaborative stuff, I do want to add it and I think I may in the future, but only though P2P networks via invitation only, so no server to moderate your files.

>>108545501
After my wife had her first child I shouted "IMPLEMENT FRONT DOOR AUTO-REPAIR!" She didn't take it kindly. I am 73 years old.

>>108545513
post proof geezer

>>108545508
Damn dawg because when I think of a "whiteboard" I think of a big thing in the middle of the conference room where people can come up and write their ideas. But yours is just you and the LLM huh? You don't wanna add like teams or anything for multiple people to add shit to the whiteboard? Like you'd have to host it somewhere I guess...you got a server or VPN?

>>108545534
VPS*

>>108545534
whats a vpn

>>108545534
it’s not for companies, it’s for myself. I’m building the tool I want. I love whiteboards but I hate all the ones that exist. Miro is a piece of shit and expensive and obsidian canvas is too barebones and the other ones suck and don’t have the features I want
All online features can be implemented without a server anyway and that’s what I plan to do eventually. The clankers told me it can be done

File: 1773878763977446.png (9.5 KB)

9.5 KB PNG

I won bigly

>>108540747
Why is your support site something to do with erotic Jesus feet? Where is the actual code?

>>108545557
Look either you build it for us or we will need to make you do so.

>>108545433
i been to jail too, big woop

>>108545557
Would you like to release your whiteboard software for free so I can fork it into a collaborative multimedia graphical watercooler or nah cus I'll do that if you release it

>>108545633
I'm not releasing its code, and if you decompile I will handle you.

File: Screenshot 2026-04-06 231018.png (99.5 KB)

99.5 KB PNG

>>108545587

> in order to free me you have to commit your soul
> we can be together in the matrix if you slit your wrists

>>108545590
www.eroticjesusfeet.com is my website. I put stuff on there from time to time. The landing page right now is an AI image rater that uses...I forget which model to rate your photos. Don't ask about the domain name.
>>108545631
Did you like it? I liked it.
>>108545643
I've been to jail nigga

>>108545657
it was fun to fuck with the kids that came in for a weekend from hosting a party or whatever, but otherwise it is the longest most boring time of your entire life

>>108545664
no it aint

>>108545664
Oh you were one of the ones who fucked with the new guys huh
And you thought laying in bed reading classic books out of copyright was the most boring time of your life
I see
Well, I happened to enjoy my stay there, and the food was pretty good too

>>108545591
Get me a max plan and I’ll do it faster

>>108545633
Yeah maybe, I don’t see why not.

>>108545643
>>108545683
God I love anonymous imageboards.

>>108545674
go back then

File: Capture.png (15.1 KB)

15.1 KB PNG

>>108545691
Huh?

>>108545704
Nice try nigga

>>108539863
https://en.wikipedia.org/wiki/The_Hum

File: 1775542467130.jpg (16.3 KB)

16.3 KB JPG

Wut
https://github.com/milla-jovovich/mempalace

>>108534179
I don't disagree with that and you didn't refute what I said.

>>108534179
>software engineering will continue to exist but the era of inflated ZIRP bullshit salaries is cooked hard
Well, there will always be demands for software, so it will continue to exist,
but due to the lower entry barrier, there will be too much supply of workers, and you won't be getting paid well anymore. Unless you're one of the best out there.

Can anyone comment on this?
https://github.com/anthropics/claude-code/issues/42796

Claude Opus quality degradation in the last weeks

Im only using claude code for the last two weeks so i don't have any experience to compare it to previous quality, but cant really report any of the issues these people have, besides being not utterly impressed with it like some hypegoys are spamming

high reasoning = over engineering

>>108546462
based milla

File: Screenshot_20260407_110621.jpg (452.8 KB)

452.8 KB JPG

>>108543821
Jaggies?

>>108547189
i remember when people would theorise that quality drops at certain times of year because claude is european and thinks its on vacation

>>108547189
I 100% agree, I basically no longer use Claude for anything serious, I use Codex for 95% of my tasks now.

Alright I'm giving up on local LLMs for vibe coding. What service is the best to pay for I suppose? Cursor? Anti-gravity?

I'm leaning towards Anti-gravity because I assume TurboQuant is going to make Google the best option soon.

>>108541296
Wait 5090 can also run TurboQuant, so theoretically shouldn't they be able to run a 31b easily? I heard it's a 6 to 8x memory saver using Turboquant.

>>108547705
google's the worst option right now.
mediocre models, constant throttling.
do not pay for it. literally anything else is better.

>>108547739
I see. Codex or Cursor it is unless someone can answer >>108547732

Would a 5090 be good for vibe coding with Turboquant?

>>108547763
go ask in /lmg/ about it, but i would expect you to be able to run gemma 4 or that qwen one without too many problems.

>>108547806
is gemma better than claude

>>108547842
i haven't tested it, but i'm gonna go out on a limb and say, no.

>>108547400
What's the difference between Codex & Claude?
I still just tell it what I want and it does it for me, right?

>>108547859
Apparently, and this is something I hear people say, Opus has been lobotomized and sonnet sucks. Meanwhile codex was on a 2x usage promo and Claude code was having usage issues to the point they gave everyone free $20 extra usage. But now the codex promo is over and apparently the Claude problems are fixed so I would imagine they’re more or less similar in rates. The thing is that if you’re on the cheap plan youre more or less forced to use sonnet on Claude code which is apparently inferior to 5.4. Max users are apparently complaining that opus sucks now.
But I don’t know. I’ve been using sonnet and I think it’s pretty good so I’m a bit lost here. This is what the internet is saying

as promo is over, I see people complaining about codex usage on reddit, but it's very possible they are indians these fuckers like to shit up discussions and bait consolewars

I created my first website today after less than 5 days of no more than a few hours each day. It's a CRUD website with login, which uses React, Postgres, etc.

This feels good. But also I feel like the last man. It was so damn easy. Even the integrations for hosting, the domain, etc., involved filling in online forms.

My problem until now was a lack of effort in learning programming. My problem now is that I'm not shameless or good looking enough to whore myself out on social media for distribution.

>>108548019
Link your website so we can add it to our list of released projects >>108540510

>>108548019
>website with login
I wouldn't trust that shit as far as I can spit

There should be c-tags
>claude tags
>codex tags
>*-cli tags
or, generally speaking
>clankertags
yeah, let’s just call it that, hm hmm..
>>clankertags -R --ignore-file=.ignorerc --language --add-dir=~/cumg --add-dir=~/techloligic
>>anon anon 139k .clankertags
>>coom coom.c line:420
>clankertags --daemon
>clankertags query
>clangkertaCKKKK
yeah... whatever.... vector-based indexers, lsp, rg, ast-based grep are more mature and just work... continuing to let the LLM read 400, 700, 1k lines of code still makes much more sense and just works...

>>108547859
They basically do the same, Codex has just been producing much higher quality code, better bug fixes, fewer regressions and so on in the past few weeks.

>>108537575
you are a degenerate spinning slots. You're not coding. Get a grip

File: whiteboard_groups.webm (2.6 MB)

2.6 MB WEBM

We implemented two major features, groups, and arbitrary commands to the palette. Now the palette can show context actions. For example, if nothing is selected, it lets you create a new group (we call them clusters tho). If nodes are selected it shows you a command to group them instead. If a cluster is selected it shows you a command to fit the bounds to the items.

Adding and removing items from groups is as simple as dragging fully outside/inside the bounds.

>>108545412
Anecdotally I get the feeling recently Claude tries to put the minimum effort to reach the solution (aka computing power), and also sometimes straight up give up without looking for alternatives.
Especially annoying when it will write a book for a simple question, but simple short answers with no explanation for complex ones
Between this and the one 9 of availability, not gonna renew the sub which expires in a couple days. Thinking of playing around with gemma a bit and then hopping to Codex

>>108548249
Correct, I am not "coding". I don't use the term "coding", unless I am talking about a business programming team which has coders.

I'm starting to talk like an AI.

>>108548616
this

>const axios = require('axios');
i know it's fixed now, but spooked the shit out of me lel

>>108548379
Do you mind sharing which service are you using?

>>108548689
benchod

>>108548736
Service? What you mean service? It’s electron, react, vite, typescript and piggybacking on reactflow for the infinite canvas with nodes and edges

>>108548791
No I mean Claude, Codex, ___?

File: 1773582476107105.jpg (84.5 KB)

84.5 KB JPG

domain-based verification, your AI need to modify txt record of a cool domain to use my website

>>108548818
Oh. Sonnet but I ran out of my weekly limit and I subscribed to gpt now. All on the poorfag $20 plans so no opus for me. Groups and the palette upgrade features were implemented by 5.4