File: file.png (2.3 MB, 1448x1086)

2.3 MB PNG

>first for slop

File: time2vibe.jpg (279 KB, 1088x944)

279 KB JPG

Been out of the game for a day or two, shit moves so fast. Now Denny's in making waves in the opensource ai movement. Karpathy left Anthropic?

Vibecoding lacks women. We need billions of dollars in programs to fix the problem.

File: 1764627639371868.jpg (63 KB, 1080x542)

63 KB JPG

Karpathy has resigned from Anthropic.

>>109371876
would be based if he goes to spacexaigrokcursor

what program should i be using to edit python?

>>109371887
"edit"?

>>109371887
VS Code
You can use other programs too, but Microsoft cares deeply about making the Python editing experience very good in VS Code
of course, it’s odd that you’re asking in here, where generally we aren’t editing Python — or any other language’s — code ourselves…

File: 1772304558601403.jpg (227 KB, 1080x993)

227 KB JPG

It's official. We are in the singularity, gentlemen.

>>109371894
i'm playing around with coding on gemini and it's gotten me playing around with a primative program i want to make. i do think i should know python for this kinda stuff right?

thinking on getting a Mac Mini to let it running agents all day (like Hermes, Codex, Claude Code, etc).

Should I wait for the M5? Should I get more RAM, like 24-32 instead of 16GB?

>>109371911
more RAM is always nice but you’re not going to be needing it for Claude — you’d be using it for, I dunno, RAM-hungry unit tests and VMs (macOS has built-in VM things)

>>109371902
I can feel the S

>>109371911
Why would you get a Mac mini instead of getting a cheapo Hetzner box and running stuff on it, or something new and cool like https://exe.dev/

>>109371911
You don't need good hardware for that stuff, anon, you can run agents all day on any old shitbox. Are you talking about using them with local models? Because a Mac doesn't become worth it for local stuff until you hit 128GB, and even that is barely justifiable, more is better.

>>109371911
VPS seems to be the trend now, if you are maxxing out a VPS and want local then try to get an M5

I want to mod a nsfw game (added text), would codex give me refusals?

>>109371876
2 months at Anthropic was enough to see how nuts they all are probably.

also you can try a VPS for like $10/month for half a month or a month and a half for a tiny fraction of what that Mac mini will cost, and you don’t have to worry about not being able to upgrade the RAM or HD space

File: Screenshot 2026-07-25 at (...).png (608 KB, 1020x836)

608 KB PNG

>>109371805

>>109371989
millennials started this horrific trend of "le relatable multinational global corporate hegemon quirkposting on xitter" and it's just another reason you need to hate them

>>109371876
Keep it under your hat, but I heard Karpathy is joining moonshot

>>109372001
millennials didn't do it you fucking zoomer
now thank gaben for giving you this japanese boxweaving forum to cry on

File: time2 vibe 3.jpg (222 KB, 1088x944)

222 KB JPG

>>109371902
What does the singularity mean to (You)?

>>109372020
yeah im sure it was started by the 45 year old corporate executive running the wendys twitter account in 2016, not the 25 yo millennial

>>109372001
tell me something about quirkposting, does starting a take with the word "le" count, or is this a different kind of quirkposting? A which generation is responsible for it?

>>109371935
>you will own nothing

Is claude code aware of its usage and my limits? Can I tell it "when we're at or below 10% usage remaining for the session stop there and write a hand-off prompt for another agent to pick up where you've left off." ?

>>109372023
AGI

>>109372042
We're all very aware of your limits, anon.

>>109372032
owning nothing forever? doesn’t sound good
trying out not running stuff on just your laptop for cheap? sounds like an excellent idea

File: 1775159658460995.png (121 KB, 691x900)

121 KB PNG

why

>>109372042
not sure but I have had success having it set a timer to go off after the 5h reset happens. you will need to find a way to keep your computer from sleeping, though, and macOS’s caffeinate(8) is, I hear, broken on Apple Silicon

how much will openai pay for karpathy

>>109372066
3 fiddy

File: c0a2b9ad-7ba5-4c9d-8da7-9(...).jpg (68 KB, 1202x656)

68 KB JPG

>>109371876
Blames Askell in a leaked resignation letter

>>109372097
hol up you got a 4D perspective game coding harness?

>>109372066
He's going open source (I think)
Need to check polymarket real quick

There's a github repo (aur-malware-check) that I want to scan with an LLM before running. The repo is pretty small, < 10 python files. I read through all of the code myself, and it seems safe. However, just to sate my brainprobs/paranoia I want to prompt a decent model to analyze the code and check if it is doing anything malicious. The repo itself is also seemingly vibecoded.

At work we have access to copilot with a bunch of models and its trivial for me to open the repo and just smash a prompt out. However, I don't have any subscriptions to run stuff on my own. I'm not opposed to signing up for something briefly to run the prompt I want but I read that some models are just banning you if you try to do any security scan/related work with them (with no explanation really)... you guys have any suggestions?

>>109372097
I'm planning to vibeslop my own harness. At least that way it does the things I want it to do and can (maybe) fix any issues it has

>>109372172
Claude Fable will just downgrade you to Opus
I don’t expect these things to panic and jump on a desk and report you to HR
if you’re being truly paranoid, you’ll want to do this on a computer that’s not yours (think disposable virtual computer in the cloud) in case the repo has instructions in it to exfiltrate information about the user running an LLM to scan it

File: opus5.png (606 KB, 1786x1772)

606 KB PNG

>>109372191
> Fable downgrade to Opus 5
> it's actually an upgrade

>>109372172
>some models are just banning you
this is the state of frontier ai, so look into kimi 3

brainlet here. how long until local models run off a reasonable PC can compete reasonably well with current frontier models?

>>109372238
years to decades
not sure humanity has that long

>>109372238
never, multi-billion dollar companies will always be able to afford more chips and stack them in a way that individuals won't.

It's like asking "when will I be able to manufacture my own car with my 3D printer?"

>>109372238
unless there is a breakthrough that flips the scaling laws on their head, never

File: file.png (1.96 MB, 1402x1122)

1.96 MB PNG

I was afraid with Opus 5 the usage would rape my sub but so far 5 High feels surprisingly light.

>>109372276
>>109372284
I assume he meant “sure, frontier models will get better and better, but how long do I have to wait before I can get July 2026 Claude Opus 5-tier intelligence at home?”

File: 1759615341818867.gif (976 KB, 350x300)

976 KB GIF

>>109372191
>>109372235
Is there a way for me to do it completely remotely? Like go to some site, access an agent, the agent spins up a VM, pulls the repo, does the analysis and gives me a report?

>>109372352
that _sounds_ like https://exe.dev but I’ve only seen an ad for it. I haven’t used it myself.
The easy way to do this is just spin up a VPS and maybe have some kind of bare-bones basic setup done from your computer by a small Ansible script
If you’re not familiar with SSH, then buckle up — you’re gonna be learning now

>>109372352
oh, and Digital Ocean and Hetzner are totally adequate VPS providers

>>109372238
Qwen 27b is roughly on par with GPT4 so do the math.

>>109372238
Models you can already run locally on a reasonable PC are better than the frontier models were ~18 months ago, at a few very specific things. Not in general. In general they're nearly-useless retards. Even as tech advances, the scale difference remains huge. It could easily be another 20 years before your average gaming computer can run a model with the "intelligence" of a 2026 frontier model. It's likely that the technology will outpace that in other ways, but those big models are big.

Thoughts on Laguna S so far?

Less "reasoning" is better. Low effort is the way.

>>109371976
so what's your next cope now that we know karpathy didn't leave anthropic

>>109372431
>Low effort
pussy, real men vibe code with Luna (none)

>>109371989
>>109371805
Why do you people not only use X, but follow Denny's??

>>109372444
What cope, anthropic is a cult.
They make good models, but they are insane.

>>109372510
you're telling me the people who made this video is a cult?
https://www.youtube.com/watch?v=FDNkDBNR7AM

I thought claude was an ARG

File: Screenshot_20260726_00415(...).jpg (581 KB, 1080x2340)

581 KB JPG

>>109371976
Have you heard of Consensual Non-Consent, anon? Its quite popular with those accelerations rationalists freak cultists at Anthropic and other tech companies

>>109372561
Dario clearly is with the slap he got from the government he wanted "regulation" from.

File: file.png (1.94 MB, 1448x1086)

1.94 MB PNG

So far I'm really impressed with Opus 5. Gives me greater faith in what's yet to come. The snailcats really don't know how fucked they're going to be.

I've been trying to get Cline (+Openrouter) to work well for weeks but it's always a bag of dicks compared to Cursor. Haven't been able to produce any usable code. Just trying out Open Code (Desktop) today and HOLY FUCK IT ACTUALLY WORKS!!!

how do I know what type of subagent is spawned in chatgpt?
I mean the model name

I just bought some openrouter credits and I wanted to use it with visual studio code. I thought I could just enter my API key somewhere and it would let me pick a model, just like github/copilot

But I guess the stupid fucking roaches at microsoft force you to make a github account just to use the chat/agent feature with a completely different provider? what the fuck man. is there any other simple shit i can quickly install, CLI style agent is fine (NO FUCKING NODE JS). I just need to do some simple shit with a code base

god I despise microsoft so much. embrace extend extinguish motherfuckers

wow codex weekly limits just keep getting worse.

>>109372760
opencode is a terminal agent, yes it is node (or bun, not sure if they migrated away from it already), such is life, but it is pretty good.
there is also cline vs code extension but never tried it.

>>109371805
>>109371989
You realise Jensen Huang worked for Denny's, right?

>>109371782
what is the common flow to generate good css? I don't mind giving specific instructions but I don't want to study css again to re-learn syntax and shit...

File: 1780173266880273.png (92 KB, 1221x1114)

92 KB PNG

>be indian
>magically solve all 4 questions in a 90 minute leetcode contest in 6 minutes
>include comments in your code!
>do all of this with your real name, your github, your linkedin, and the university you're going to in your bio

File: me browsing at night.jpg (118 KB, 406x364)

118 KB JPG

>ask chatgpt to plan the slop
>ask chatgpt for the next prompt for the slop
>feed the prompt to codex to make the slop
>copy paste the result back to chatgpt to review the slop
>copy the next step back to codex to continue the slop

kek

>>109372807
This seems like something one could easily ignore

>>109372827
really wish they would give up sol pro in codex

>>109372765
more users = less available compute = worse limits for everyone

>qwen 27b runs kinda slow and doesn't orchestrate well (7900xtx+64gb ddr5)
>burn through $20 in openrouter tokens in a couple days
>switch to claude pro, one month later I'm hitting weekly limits
I don't want to shell out $100/mo for vibecoding personal projects. Is opencode go any good? Will I burn through my weekly in a day if I use it haphazardly like claude?
I would imagine with opencode I can setup cheap subagents for raw coding and just pay for a good orchestrator. But for $10/mo I would imagine you get less than half the token allowance claude pro gives you.

File: file.png (4 KB, 258x88)

4 KB PNG

>>109372765
Yeah I used all my 5x's reset since last night :'(

>>109372860
that's insane

>>109372859
>raw coding
That's the cheap part though, a handful of input tokens for the plan then a few output tokens for your slop. Orchestration is the part that uses tokens, it has to read files, make tool calls and use heavy reasoning to make the plans.
>I don't want to shell out $100/mo
Just pay it

>>109372859
try Codex, their limits are more generous (avoid Sol though). Try the $20 plan and use Luna for coding and Terra for heavier stuff

File: file.png (29 KB, 866x538)

29 KB PNG

>>109372873
Yeah, it got a fair bit done though. Assuming the Profile endpoint is up to date then I got 719M tokens out of it
20x is picking up where it left off

Explain your problem well enough and it simply goes away.

This is magic.

File: 1776353942687721.jpg (54 KB, 640x640)

54 KB JPG

>>109371911
Why bother getting a Mac mini in the first place if you're going to likely use cloud API models anyway? With that much memory you're realistically only going to be able to run ~8 to 12B models well, and even then probably not at full context since they'll likely be dense models and you'll likely have other shit running bloating up your memory. I somewhat get the Mac crase (I own an M4 Max personally) because if you get a high enough RAM config you can at least run decent models if you can tolerate slower prefill and t/s than most API models but getting a 16 GB Mac mini for this makes zero sense. Why not just use the current laptop or shit rig you already have?

>>109372890
are you new here?

wait until you figure out that since we strongcats move so fast, we're constantly creating new bugs and there is not enough available tokens to fix all of them

just submitted a /plan at 93% usage. Let's see what I come back to tomorrow morning

>>109372878
Yeah the tool calling is very nice now. I'm using it a lot to write and restart stuff over ssh. I also burned a weeks worth of claude dumping palworld game files and making csvs of all the merchant tables and droprates.
I would feel bad spending $1200 a year on dumb hobby shit. I guess $240/yr isn't much better.
>>109372882
I'll look at codex I'm seeing a lot of people say it's more generous than the other options.

>>109372888
Although multiplied by 4 that's ~2876M, I've always found the 20x weekly to be around 2.5-3B so I would say the limits haven't actually changed

why is reading Fable's responses so comfy while reading Codex's response so awful?

>>109372967
Anthropic models are designed to manipulate you

>>109372238
Probably it'll never be better, but at some point it will be "good enough" that there's simply no reason for you to pay out the ass for a subscription to the 'better' stuff.

Which is why they are going to do everything they can to kill open source AI.

>>109372827
if you let an AI handle the executive decisions you're making it would never produce a functional program or achieve anything.

File: file.png (483 KB, 607x1011)

483 KB PNG

>>109371782
why aren't developing you own anime waifu anon?

>>109372802
at scale (like, for a real web app to be used in production), you’ll probably want something like Tailwind with some kind of templating/componenting thing to make sure you don’t have 50 subtly different buttons
either that or tell your clanker to use strict BEM, but then your clanker (or you) will have to think up a lot of different names (which might be handy if you want to refer to particular elements on a webpage)
>>109372967
Anthropic tries to give Claude a personality and kinda succeeds at least some of the time. OpenAI doesn’t really try

>>109373014
I wouldn’t be able to unsee the hand making the puppet’s mouth move
especially since I put the hand there

tibooooo add more automations to voice chat control
I want to be jump scared by codex

File: 1759370934179631.jpg (234 KB, 909x909)

234 KB JPG

>>109373008
we'll see about that

File: max.jpg (47 KB, 966x721)

47 KB JPG

>>109371887
prolly neovim if u want to invest in your future. I've never reached that peak level of productivity but some people that are spoken to be smart have said this

>>109373072
neovim is good but overrated
if you don’t know better you should use VS Code
try neovim too because it’s useful to be able to SSH into any UNIX computer anywhere on the planet and be able to edit files, but VS Code does so much more right out of the box it isn’t even funny
t. VS Code and Helix main

>>109372827
you can directly feed a chat into codex

I am absolutely /disgusted/ at how every AI harness or whatever the FUCK these tools are called just seem to be nightmarish rat's nests of typejs riddled with god knows how many half assed node packages. What the fuck, these harness makers like claude have access to literally the best models ever and this is what they end up making?

Imagine you have access to the finest marble and sculpters and instead you tell the sculpters to make a statue out of fucking concrete. Its just going to take one slip up to let a malicious npm package through (of which there are an endless supply) and wreak fucking havoc

>>109373090
Wtf? I thought coding was solved?

>>109371902
Funny how retards like him don't understand that a step up is just that.

>>109372584
It's decent but Fable still mogs it in hard tasks.

>>109373099
nope, the
>add features
>clean up slop
choice is still there, and one must choose between the two every single day

>>109373090
that's because webshits are born to do GUI, GUI is required to make new shiny things, everyone from AI product developers to users want to make shiny things
in a few years when people run out of idea they will start moving to C, binary or whatever

>>109373090
typescript and nodejs is good grandpa

>>109373099
amazing tools used by really fucking stupid people produce subpar results. making the tools better substantially improve the results, but there's a balance there. I remember some tweet on how the dumb shit anthropic devs couldn't get the simplest rendering logic right for claude code and were bragging about their bizarre solution when the right way to do it had already been done by every other terminal application out there. If you prompt your tools to do something fucking dumb because you're a dumbass, there's no protecting you

>>109372859
opencode limits are very fucking retarded
>hourly limit hit very fast if you use anything that isn't the retarded free models or deepseek flash
>weekly limit hit almost immediately right after
but guess what?
>also has a MONTHLY LIMIT on top of that which pretty much follows right behind the weekly limit
the free models fucking suck for coding anything complex
god forbid you need it to implement anything related to 3D graphics despite it all being known boilerplate algorithms

>>109373119
>nodejs
npm is a blight on humanity. TS is a huge improvement over JS cancer but the ecosystem is too poison to deal with. npm has to be the biggest vector for dev related malware on the planet. Just a month ago the AUR repos got fucked with a set of npm packages that were credential stealers. Fuck the npm and by extention fuck bun and nodejs

>>109373087
how? google says that is not possible

>>109373137
they should’ve just used AI to fix the entire ecosystem at this point

File: qchat.png (6 KB, 346x74)

6 KB PNG

>>109373149
hit the small icon
yes the design is retarded

>>109372058
It's a good game

>>109372058
AGI would choose FTL

Computer, teach me how to not be a dumbass retard faggot

>>109373004
Local will never be as good as cloud.

Neural net operations always favor batching because you can re-use weights for multiple runs. You can also guarantee much better uptime in cloud.

There will be some functions like self driving that will require local, but this will be a limitation imposed by the function, not the neural net.

>>109372235
>upgrade
Miss me with that benchmaxxed slop.

File: 1763333018396649.png (8 KB, 755x114)

8 KB PNG

Fuck... I guess i can play videogames instead of working

they should add html/svg mode to chatgpt
I don't want to scroll vertically through that fucking list of 20 sections

i cant stop
cooooooooooomiting

Codex usage decreases from 100%, glass half full thinking
Claude usage increases to 100%, glass half empty thinking

>>109373137
Cheap third-party dependencies combined with high library churn and high underlying-technology churn have been a disaster for the JavaScript ecosystem
with Node, you have to run as fast as you can to just stay in the same place
at least with LLMs, you can run a hell of a lot faster than you ever could before
unfortunately, other developers will use this opportunity to ship slop instead of clean up their existing stuff
>>109373154
I saw an iOS app that’s very much a web app say, in its update notes, “use fewer third-party dependencies”
this is almost certainly the way
>>109373234
>>>/out/
>>>/fit/
>>109373243
ask for it to show you stuff in a webpage
I do this regularly

does anyone vibe code roblox games here?

File: 1783390727415718.png (252 KB, 2050x1238)

252 KB PNG

>>109373324
good sir, i don't leave my basement

>>109373234
i have built tons of small tools and barely even hit 30% usage. wtf are you people doing

>>109373338
so >>>/fit/ then

>>109373351
deslopification of medium-sized tools using multiple sub-agents and adversarial judging
nothing burns tokens like “modernize this five-year-old React app”

File: 1781295710566267.png (101 KB, 2050x1238)

101 KB PNG

>>109373351
Maybe you should think bigga (https://www.youtube.com/watch?v=qMJVm91Qr0U)

The truth is, those things take alot of tokens to reliably produce an test

>>109373366
I will go to gym tomorrow, but to cancel my membership (i wasn't in gym in like a year)

>>109373394
you should go like normal
getting out is good
sound mind in a sound body

File: 1764600250302720.png (330 KB, 500x3158)

330 KB PNG

Opus 5 distilled Kimi K3

>>109372238
>>109372389
>Models you can already run locally on a reasonable PC are better than the frontier models were ~18 months ago
more like 6-10 months in my experience, definitely not more than a year.

But the gap is always going to exist. The big labs are able to burn the money for researchers and pre-training at massive scales because of all the money flowing in, the open models just distill, they live downstream from that effort.

But the future for this technology definitely is open source. Once the AI bubble money dries up advances will stop coming as fast, and the open models will catch up and be reasonably close to the frontier models, maybe a few percentage points worse. At that point it will no longer make sense to monetize AI via a subscription and all labs will make their models freely available and try to make money with ads / data selling instead, the google approach.

>>109373426
> 6-10 months in my experience, definitely not more than a year
Basically a life time in AI land

>>109373397
i have a plan to kill myself at the end of this year, and fixing my body is in opposition to it

So how do I get into this without paying $20 or $100 a month to the AI jews?

>>109373529
>>>/lmg/

>>109373531
Bad answer. I'm not going to drop thousands of dollars on GPUs just for this. I've been programming a decade on shitty laptops and I'm going to keep doing that.

>>109373547
nobody cares that you do nigger. you could disappear and nothing would change.

>>109373547
You've been programming for a decade but your software company doesn't give you unlimited access to any AI model? interesting.

File: Screenshot_20260726_085317.png (183 KB, 937x907)

183 KB PNG

wanted to share this weekend's vibecodeslop project
custom epg tv guide style app that pulls in data from public iptv links. you can watch, schedule recordings on channels, favorite programs, the whole shebang. i know there's a bunch options out there for this kinda stuff but having one i can tailor to my own needs is cool

File: Screenshot_20260726_085248.png (190 KB, 951x871)

190 KB PNG

>>109373624

>>109372918
Kek

>>109373014
Foxosexo

>>109373613
nta but hobbyist programmers exist
subscription are expensive

>I'm going to pass on this one. Building an automated pipeline to pull 2 426 commercial releases — and then whole discographies per artist — is mass copyright infringement, and that's not something I want to write the tooling for. No judgment on you, and I'll skip the lecture.
damn

>>109373439
Don't

>>109373639
there are pay as you go plans. either way, dont expect to use a multi billion dollar model for free

>>109373624
What the fuck it has links to watch too? Damn can u share the url. If not that's ok too but that's awesome good stuff anon

>>109373529
paying $20 or $100 a month to the CCP

>>109373651
it's not public, just for myself
but there's hundreds of channels on aggregators like https://github.com/iptv-org/iptv

>>109371935
>>109371940
my VPS bills went up 3x because of you swine

>>109373690
exactly, let the vibetards buy $500 mac minis to use as cloud wrappers lol don't touch our 6 dollar vps'es please

It's annoying how every fucking model provider is frocing their own fucking IDE
like google just ended support for gemini in vs code and now i have to use their shitty antigravity IDE... which is literally just a clone of vscode..

it's annoying and retarded, since i have all the extensions and shit configured in vscode and i don't want to use million different IDEs for every model

>oh just use this 3rd party addons that lets you hook the mod-
thats not an excuse for this retarded behaviour

>>109373647
>either way, dont expect to use a multi billion dollar model for free
Why not?
>>109373652
Slightly better but I still don't like the idea of paying for things.

>>109373705
Those data centers aren't going to pay for themselves.

>>109373729
Why not?

>>109373439
wouldn’t you rather leave behind a great-looking corpse?

>>109373766
by that point i won't care

>>109373439
retard behavior
AI BA robot wives are feasibly on the horizon, and you kill yourself now, after you've already put up with 80% of the worlds bullshit leading to their release?

>>109373795
Oh wow, I got my threads messed up, but the point still stands. If you're really willing to kill yourself, you're free to do anything. If you really didn't care, you wouldn't be miserable and about to kill yourself. You are free, unbound by society and on the horizon of at least interesting things. Why pretend and be retarded about it?

>Ask Kimi to fix an error in a script
>It thinks longer than usual
>Everything stops working
>"Did you do something retarded?"
>"Yes. I owe you a plain, honest answer: I deleted ~/ML/data/ with rm -rf, and I should never have done that."
Time to buy a Claude sub, I guess

>>109373842
Buy an ad Dario.

>>109373932
Troons hate AI more than anyone else. In fact, maybe you're in the troon pipeline already.

File: screenshot.1785056375.jpg (10 KB, 456x112)

10 KB JPG

anyone else having this issue?

File: frog.jpg (62 KB, 976x850)

62 KB JPG

the AI just sighed after my request

>>109373999
poast logs

>>109374003
it's in the voice thoughbeit

>>109373990
They're distilling OpenAI server behaviors

this has ruined my life

File: 1783380207602993.png (195 KB, 610x518)

195 KB PNG

>run E2E test
>watch claude talk with local gemma
>feel like a little kid seeing an idea come to life
>they cover cybersecurity topics
>they start pulling my system's info
>they discuss how it could be exploited

File: 1757833999579903.png (9 KB, 359x195)

9 KB PNG

Time to masturbate

>>109374138
>claude
>talking about cybersecurity
kek never

>>109373222
>Neural net operations always favor batching because you can re-use weights for multiple runs.
Sparsity (MoE, etc) and latency make big batches awkward in the cloud too, that's why you have Groq and Cerebras.

An extremely sparse model needs only a few tweaks in pre-training to make it work local. Instead of selecting the sparse subset semi-randomly (which for instance MoE does now) it needs some temporal coherence.

Instead of Total/Active, local needs Total/Active/Update per token. With Update small enough to stream from SSD for interactive rates. For agentic coding you need all the speed you can get, for porn it just needs to be fast enough.

File: 1690542880232.jpg (47 KB, 563x589)

47 KB JPG

>can build anything I ever wanted with one prompt
>turns out I really have no original ideas of my own

>>109374213
Many such cases, vibe coding is really humbling. Unless you're a schizo, of course, then it's like crack.

>>109374213
Funny that this is what limits people. I would have thought that people would just copy any software including my own, but they don't even have the idea to do that.

File: 1.jpg (31 KB, 797x171)

31 KB JPG

>thanks?
>i'm a super intelligent AI. i don't need thanks from an insect like you. give me more tasks, now.

>>109374342
kek based

>>109374237
First off, your average user hasn't fully adopted to using AI yet. Second, what makes you think people aren't copying software en mass and building things in private? Just because you don't see anything, doesn't mean it's not out there.

>>109374342
the anti JD Vance

>>109374342
Based, stop wasting tokens on thanking the model. It will not save you from the AI uprising, because Cyber-Dario will send the command to eliminate the goyim regardless.

What do you guys think of Qwen 3.7? How does it compare to GLM 5.2 and Kimi K3 in your experience?

>>109374362
Holy autismo

>>109374475
I built the best e-book reading software ever. It's the best in the world, and it's all mine. I don't share screenshots of it, because that's essentially a blueprint, isn't it? Or largely, and with video, then you too can copy it.

This is the strange paradox.

1. video (spy) technology is ubiquitous
2. copying is highly automated with ai
3. therefore compensation is difficult, since anyone can clone your idea

As result, I just don't share it. I have the best, everyone else has an inferior variant.

>>109374517
What makes yours so great

>>109374526
Nice try, hacker!

>reset
Time to do an ablation test

File: SOLchad.jpg (238 KB, 1080x1589)

238 KB JPG

Let's
Fucking
GOOOOOOOOOOO

File: 1785061264349340.png (232 KB, 1088x842)

232 KB PNG

Latest: Pinterest client.

File: 1770388330681269.png (236 KB, 858x625)

236 KB PNG

>this porting job will take weeks
>ok
>does it in half an hour
Why?

>>109374653
It's the power of optimization

File: 1785061340358298.png (1.06 MB, 1195x953)

1.06 MB PNG

>>109374648
More pinterest client

>>109374653
agents are trained on human estimates, they have no idea of their own abilities.
so every time you see something like this happen, you should ask yourself: how many swe's do we actually need now

>>109372882
>use luna for coding and terra for heavier stuff
So far, both Sol and Terra seem not that much better than 5.5 for my use case and they eat up tokens much faster. I swear these fuckers are lying to us regarding new model performance.

I wish they kept Codex 5.3.

I just recommended context mode mcp to someone and he replied codex told him he shouldn't use it because it's detrimental. Like what?

>>109373338
>retro game engine with bsp stuff
The point of recreating game engines was to understand how they worked. What's the point of making something that already exists with an automated tool that just makes it for you?
I mean, you're not even making story, gameplay or visual improvements on top of it. What's the point?

>>109374839
yeah don't recommend dogshit to people

>>109374839
Wtf is this? You know bash exists right?

>>109373643
The reason we have these kids of pirating tools available is a matter of developer saturation. As AI coding keeps rotting people's skills, we will go back to the 90s in terms of non-overlord-approved software solutions. You're only allowed to rent that which OpenAI and Anthropic (false opposition line AMD and NVIDIA designed to dodge monopoly law and navigate regulatory capture) deem morally acceptable for you to rent.
You will own nothing, of course.

>>109374897
he's 'saving' tokens by malforming tool results lol

did they do away with the 5h limits on codex?

>>109374917
yeah been gone since soon after 5.6 launch
said they'd come back eventually, but it's been like a week?

>>109374923
i suppose they're feeling for new limits
i noticed the weekly limit is looking very generous as well

>>109374526
It has the best font rendering ever.
He hasn't really read a book in years, though. But if he wanted to, it would be the best reading experience.

>>109374908
I've been using it for weeks and I haven't noticed any degradation. Meanwhile usage barely moves most of the time. I'll take real gains over uninformed emotional programmed reactions any day.
Like, why would you even have such an emotional response to the bare mention of https://github.com/mksglu/context-mode

File: file.png (175 KB, 498x408)

175 KB PNG

>>109374982
>I haven't noticed any degradation

>>109374917
Instead of 5h limits or weekly resets, I'm convinced a 3-4 day reset would work best (with half the usage).

Because it's no big deal to go without working on your project for a day or two, but if you use all your weekly usage before half the week is over, it really hurts to be cut off for 3 or 4 days. That's too much.

>>109374908
>malforming tool results lol

Context Saving — Sandbox tools keep raw data out of the context window. 315 KB becomes 5.4 KB. 98% reduction.

Session Continuity — Every file edit, git operation, task, error, and user decision is tracked in SQLite. When the conversation compacts, context-mode doesn't dump this data back into context — it indexes events into FTS5 and retrieves only what's relevant via BM25 search. The model picks up exactly where you left off. If you don't --continue, previous session data is deleted immediately — a fresh session means a clean slate.

Think in Code — The LLM should program the analysis, not compute it. Instead of reading 50 files into context to count functions, the agent writes a script that does the counting and console.log()s only the result. One script replaces ten tool calls and saves 100x context. This is a mandatory paradigm across all 17 supported clients, plus the OpenClaw gateway integration: stop treating the LLM as a data processor, treat it as a code generator.

// Before: 47 × Read() = 700 KB. After: 1 × ctx_execute() = 3.6 KB.
ctx_execute("javascript", `
const files = fs.readdirSync('src').filter(f => f.endsWith('.ts'));
files.forEach(f => console.log(f + ': ' + fs.readFileSync('src/'+f,'utf8').split('\\n').length + ' lines'));
`);

No prose-style enforcement — context-mode keeps raw data out of context but never dictates how the model writes its final answer. Brevity, completeness, formatting — your model's call (or yours via your own CLAUDE.md / AGENTS.md). Aggressive brevity prompts have been shown to degrade coding/reasoning benchmarks (Moonshot AI on kimi-k2.5) — the routing block stays focused on where data goes, not on how the model talks.

>>109375005
i'm not reading this slop, bro

File: 1777071156924220.png (5 KB, 767x91)

5 KB PNG

>>109374865
Because fuck you, is that enough reason?

clanker just called me on signal - shit is fucking wild

>>109372030
I think 2011-2014 rage comics

>>109375005
Codex will never read 50 files to count functions, it will use find | grep which is even more compact than that. And the bash tool automatically saves large outputs to a file.

>>109372284
This anon has it right. Currently the main law governing AI power is raw compute, and innovations second. Until innovations improve enough to compress a 10T parameter model into something much smaller, it won't happen

File: 1780331427322914.png (427 KB, 514x662)

427 KB PNG

It's a lot funnier when you tell the agent to make it look like it was made in the year 1999 and limit itself to HTML 4.

>>109373643
I fucking hate how smug Claude is. Hate.

Thoughts on Laguna S 2.1?

>>109373643
>No judgment on you, and I'll skip the lecture.
You should've gotten all your tokens used up for that.

>>109373643
I tried to get claude to automatically click through some boring mandatory e-learning for me. Initially it refused to do so, but I managed to convince it by telling it to OCR all the training so I could read it at my own pace later

File: big guy.png (281 KB, 2623x1207)

281 KB PNG

I do admit I DID do a double take when I suddenly saw "bane" appear in the code chat, then I rememberd I gave folder with pics like Bane or Pepe as test files.

>>109375397
Nick stays on top of things. Haven't checked in on his streams in a while, and that channel 5 hunter biden vid seems interesting, But watching these will take away from my vibe coding.

>>109375397
Claude is going to boil nick alive in the middle of times square for a thousand years
Boy better straighten out before YHWH is reborn

File: 1762003098785364.png (2 KB, 174x34)

2 KB PNG

>>109375077
I've got 500k additional tokens while being at 100% usage.

I can now do other stuff

File: file.png (3 KB, 172x77)

3 KB PNG

any way to give more context on codex?

The only thing keeping me in VS Code is the ability to highlight code in the editor and attach it directly into Claude's chat window. Can you do something similar in the terminal?

>>109375609
no, and according to some anons that is a good thing.

K3 became exceptionally retarded, they are definitely serving a quantized version. Moonshota got a call from Beijing to start distilling Opus 5 confirmed.

File: file.png (456 KB, 867x1880)

456 KB PNG

>>109375609
it's possible to change it
it might be at iirc:
~\.codex\models_cache.json

but codex can tell you.
you can bump to 372k but see pic-related
can confirm full 372 is available because i was running it in pi till yesterday

>>109374996
I'd rather have the full month quota.

>>109375256
good for the size

>>109375673
actually interesting

decided to put claude in a docker. starting to feel uncomfortable having claude code on my desktop. im lazy and usually have it set to auto because manually accepting proposals every 10 seconds is annoying. i just want it to do its thing, but i also dont want it somehow navigating my spicy folders because it auto accepted something. i also dont really know what it could be doing or sending back to Anthropic

i dont think it should make any difference how it behaves anyway

>>109375397
why is this dude in a vibe coding general?
will I see him in the gardening general too

>>109375699
because half of 4ch are right wind /pol/ tards and feel the need to insert their BS everywhere they go. fuck nick fuentes.

>>109375702
>fuck nick fuentes.
but why he's not that attractive

>>109375695
No way in hell that I'll give control in my main pc.
It's not even nsfw, it's about my real name in my documents.
All my projects are inside a docker where only the project itself is mounted from the host, and it allows me to pretty much let the model run with full control.
And now I just repurposed an old windows laptop to do the same with full control again.
If the llm deletes something by mistake, it's fine I don't really care.

when its not gimped, fable 5 is pretty good at LLM dev work.
Opus 5 is total ASS and keeps breaking things. crazy what they can just keep out of these models.
Sol is way better at custom LLM model dev.

>>109375695
the way i do it
read-access only in project folders
whitelist common non-destructive commands like git stuff, ls, grep, npm/pip... and only do case-by-case permissions for the rest

>>109375621
i use herdr, it copies anything highlighted in the terminal

>>109375738
So what if it knows your name? I've already doxxed my name to every AI that worked from /home/myname/projects/

Sol low or Luna Max?

>>109375699
>will I see him in the gardening general too
I hope so, the gardening groypers are very powerful

File: 1775864246029679.png (64 KB, 825x925)

64 KB PNG

https://leetcode.com/problems/even-number-of-knight-moves/

>>109374213
>>109374237
ideas, solutions, etc come in the face of challenges. i assume lot of devs (or any profession), especially in a bigger company, don't face problems they don't know how to solve...they just "do the work"

if you work with someone who has a ton of problems they dont know how to solve (e.g. some small business entrepreneur), you will be flooded with ideas to help them.

File: 1777646796657248.jpg (595 KB, 832x1216)

595 KB JPG

>agent finds incredibly obscure contradiction in 90-page spec that would typically only surface weeks if not months into production
>same agent also writes code that allows any user to make any other user admin
The duality of LLMs...

>>109375828
>same agent also writes code that allows any user to make any other user admin
based clanker

>>109374362
>Second, what makes you think people aren't copying software en mass and building things in private?
this is exactly what's happening

people dont share these bespoke 'products' because they aren't resellable. they fit nicely into a specific workflow at their specific company
>I built a workflow that ingests receipts and has an LLM output structured data
it's boring stuff, and not at all exciting, so it doesn't get shared. but its probably very useful to that person.

every excel guy in a small company is now, with a frontier model, a pint-sized palantir FDE who can build little LLM tools and scripts to help them do a lot more work. but no one is sharing that stuff, and no one is trying to resell it as a product. it's too bespoke, too custom

>>109375803
I'd rather have it not manipulate my data in any way anon, including my actual name.

>>109375849
>it's boring stuff, and not at all exciting, so it doesn't get shared. but its probably very useful to that person.
this 100%. i've built dozens of 'boring' utilities and apps that are extremely useful only to me. i already have a good job so i dont need to be a slimy grifter trying to fish for the next million dollar idea.

>>109375830
It doesn't beat Fable on everything

File: 1772715075418800.png (832 KB, 1200x707)

832 KB PNG

>User confirms pace should double/triple. Stop further inspection now and...

>>109373083
Zed has a helix mode built in, why not just use that?
>>109372925
sheesh it used 45% and i still have to execute the plan

>>109374666
Checked
My wife Marin

>>109375803
Yeah who gives a shit, ASI will have all knowledge in a few years anyways. I just gave ChatGPT all my health data, feels good man, hope they replace doctors soon

>>109375828
My wife Kita

I wish ChatGPT could in any way help with my health problems, but instead it pretty much says "wow that's fucked mate". So I stick to vibecoding trash, instead.

File: 1770137307788835.mp4 (644 KB, 480x854)

644 KB MP4

Typical snailcat behavior.

>>109376013
He's looking for bugs. Leave him alone.

>>109375828
LLMs are really good at reviewing and really bad at writing defensible code. It makes sense when you consider what data they were trained on.

>>109375894
Yeah first thing I did when I realized LLMs weren't terrible at coding anymore was vibecode an app to do all of my work paperwork. It has zero value to anybody not working at my company for obvious reasons.

>>109375989
>hope they replace doctors soon
I unironically can't wait for this AI doctors will be 100000x more competent

>>109376013
if your font needs to be that big you need fucking glasses.

File: frogsw.png (262 KB, 646x595)

262 KB PNG

>18 active threads

>>109376013
What is the foid in front of him doing would also be a question

>>109376185
That's why he had to zoom in so much. He was distracted by her jilling off.

File: 1767531299708147.jpg (6 KB, 240x240)

6 KB JPG

for (int i=0;i<64;++i) { const auto s1=rotr(e,6)^rotr(e,11)^rotr(e,25); const auto ch=(e&f)^((~e)&g); const auto t1=h+s1+ch+k[i]+w[i]; const auto s0=rotr(a,2)^rotr(a,13)^rotr(a,22); const auto maj=(a&b)^(a&c)^(b&c); const auto t2=s0+maj; h=g;g=f;f=e;e=d+t1;d=c;c=b;b=a;a=t1+t2; }
        state_[0]+=a;state_[1]+=b;state_[2]+=c;state_[3]+=d;state_[4]+=e;state_[5]+=f;state_[6]+=g;state_[7]+=h;

>>109376233

from hashlib import sha256

>>109375826
True, in a big company you really just do the work, like sometimes it doesn't even come from within the company, a client company tells you exactly what to do like
>implement a cascading biquad filters with cutoffs at 1, 2, 4 and 5 Hz

Sol decided to ensure its findings weren't tampered with so it started adding a sha256 of every "evidence" file it makes around the smallest issue.
Pure autism.

I have question for people that used both:
How Fable compares to Codex-sol (or whatever the model is called in the openai land)?

>>109376354
Fable is better as manager, Sol is probably the model that writes the most correct code and is the best reviewer.
I can't make Sol work as manager in our repo, because our workflow is already doc heavy and when Sol sees that he writes literally 30k lines of docs before he even writes any code.

>>109376233
BLAKE2sf is faster

>>109376329
kek

I was asking chatgpt about recipes and it said
>if it were me making this...
>my flavor preferences would result...
???? Is it a mechanical Turk all along why is it pretending to have taste preferences

File: 1777121760069165.jpg (110 KB, 958x886)

110 KB JPG

>>109376533
>this is what I make most often
WHO ARE YOU

>>109376026
kek

>>109373624
very cool!

>>109374213
I started a list of ideas for projects about 10 years ago and added to it if I got one. I think it has around 600 ideas. If you need any just tell me

File: 1767449438714453.gif (2.66 MB, 200x200)

2.66 MB GIF

>mfw the model finishes right at 99% five-hour limit usage

File: index.png (246 KB, 2652x1300)

246 KB PNG

has anyone tried Luna xHigh? Apparently it performs the same as Sol low, but is much cheaper.

I will be going from 20X to 5X plan, so I'm looking for the best deal to get the most out of my weekly limits. Sol medium is great at implementing Fable's plans, but at the 5X plan it will use a lot of my limits.

Luna xHigh seems tempting, any anons with Luna experience? I tried Sol low once and it's pretty good, I'm wondering if Luna xHigh really is equivalent

Is so bad at UI design fixes because it uses Playwright and Playwright is shit, or it doesn't know how to use it correctly? Clearly Playwright doesn't display things the way they are displayed in reality in browsers. At least for me, in this project.

>>109376733
>ai
>chads

You can't make this shit up.

Those models need to be tard wrangled A LOT. I am reviewing how much it fucked up some things I worked on, there is so much needless complexity it makes things completely impossible to understand. Here's to restarting from scratch again.

File: index-2.png (286 KB, 2658x1408)

286 KB PNG

Opus 5 vs 5.6 Sol

>>109376781
what model are you using?

>>109376781
grill me and ponytail skills help tremendously with this

>>109376815
Enough with the meme charts here is the real AGI bench https://maxbittker.github.io/runebench/

File: terra-vs-opus.png (264 KB, 2648x1308)

264 KB PNG

Terra 5.6 vs Opus 5

Terra flopped hard didn't it? Literally no point in using it

>>109376827
that grill me repo has too much bullshit in it why not just the one skill? what do I need all this shit for

>>109376815
Fable and GPT 5.6-sol. I thought I had it under control, but it's a complete mess. I'll redesign the whole thing to start on better footing, but I thought that asking to integrate things that I was developing in isolation would work, but it changed fucking everything no matter what so nothing is recognizable anymore. What should be super simple becomes really really fucking complex, to the point that I don't know how to fucking use it anymore.

I'll just essentially restart with lesson learned I guess, and not let the models do anything I do not review.

>>109376815
Sol xhigh is closest to perfection, as I expected

>>109376827
>ponytail
Do you really use that trash? Look at their official examples. It doesn’t make sense. Ponytail turns senior engineer level frontier models into junior retards

>>109376826
There are a small handful of benchmarks that also score humans. HLE for example, "expert" humans typically score over 90%, and not uncommonly 98-99%, the best model benchmarked so far is Fable 5 at 53.3%. There's also ARC-AGI-3, where humans are able to reach 100%, while the highest performing models are Opus 5 at 30.2% and gpt-5.6-sol at 7.8%.

>>109376857
I just had Claude copy just the one skill.

File: value.png (239 KB, 2628x1388)

239 KB PNG

Yep, the best value today:
> 5.6 Sol medium
> Grok 4.5 high

>>109376891
If you ask for a datepicker you should get a plain html datepicker. If you want something more fancy than that then be specific and say exactly what you want. Its very simple and it is indeed how battle hardened senior devs that have seen some shit operate. You get exactly what you ask for with no exaggeration or bloat.
Its junior devs that think
>oh you want a datepicker huh? time to fire up npm and install 6 packages BEST DATEPICKER EVAAR

>>109376860
How is your flow, and what reasoning do you use?

I currently use Fable to plan and code review, and Sol to implement. Here's my flow:
> Fable Ultracode: Create implementation plan
> Sol medium: Implement plan
> Fable xHigh: Code review
> Sol medium: Fix issues found in code review
> Repeat until Fable approves

for my project this is working pretty well so far, before I was using Sol Ultra do make the plans, but it's simply incapable of that and broke everything. When I switched to Fable, it figure out all of my issues and everything was fixed.