>>109312683
Coding agents suuuuuuuuuuuuuck !

>>109312689
this but the opposite

>>109312689
(my penis)

Unreal Engine developer here.
How usable are Claude Code and Codex for helping with Unreal Engine C++? And has anyone here tried the new Unreal MCP tools inside the engine? I understand they consume a shitload of tokens though.

>>109312728
dunno

>>109312728
yeah

File: 65737455634543254325432.png (388 KB, 807x604)

388 KB PNG

Quick, tell me about this new Kimi K3 thing. How is it? Is it on par with Fable?

>>109312749
see >>109312737

>>109312749
yeah its good
soon coming to your desktop in a quantized two bit model

Why does no one use Cursor anymore? It was everywhere like maybe 1-2 years ago

>>109312777
we don't look at code anymore

>>109312777
Back then models were retarded and you had to work together with them. Now they work instead of you, so people prefer the simpler cli solutions where you just go "/goal get me a gf, make no mistakes" and then check the result next day.

>>109312777
we don't look at code anymore

have sex with the chicken

best vibecoding stream
https://www.youtube.com/watch?v=V6mkBjL26ps

I rarely used AI untill recently, Ive used claude for some coding stuff before and its gone ok but I just had gemini results in a google search, tell it im wanting a python script for a certain api and it shits it out nearly instantly. Kinda insane how frictonless things have gotten for simple tasks...

>Models are trained on old jeeted TODO apps and follow retarded inefficient patterns
>Spend hours hand-crafting Swift 6.5 skill, explaining how lifetimes, references and ownership works, as well as the modern SwiftUI practices
>Start a project that will clearly benefit from it
>Clanker sets the target to v14, locking out all the features
Man, there's just always one more thing you forgot to mention, and everything falls apart

I find it kind of sad that almost all serious careers are safe from AI, but programming isn't one of those, it got killed by vibe coding completely

File: 1771420421391323.png (154 KB, 590x350)

154 KB PNG

>trying to find business partners
>first guy leads me in circles despite being aligned on vision/strategy
>he's also unreliable so I'm 95% certain I'm going to back out on him
>now asking my friends with compsci degrees
>unironic potential to earn 700k-1M/yr
>see my commits
>whoa, you're working a lot man
You know, the crazy thing is that the free version of my app is already released and I have thousands of users, positive reviews, and hundreds in donations. Why is it so hard to find a young and motivated cofounder? I'm trying really hard. I feel like I have everything that would look desirable to someone but all of my potential partners always seem to be only half interested.

I just want someone who will work with me. I'm not even opposed to going 50/50 on equity if they have the right skill set and prove themselves useful.

FUCK. I really don't know what more people need, perhaps it's better I just continue on my own.

>>109312983
ehh i wouldnt be so sure

>>109312991
fuck off grifter

>>109312683
Do you need liquid cooling to vibe code?

>>109313012
unironically ive been thinking about buying an used server and dumping it in mineral oil or some other coolant
i want to run llms for cheap without waking up the neighbours

>>109312983
>almost all serious careers are safe from AI
i'd wager almost all knowledge work could get significant productivity boosts right now
enough to lead to substantial reductions in staff, but the people who own those businesses just don't know it yet
i personally know someone who works in project management who refuses to automate work that he admits can be automated because it'll lead to layoffs

Can anyone give me a quick rundown on how to vibe code the web frontend? Particularly in the case when you already have a design in Figma or somewhere. Models seem to make up some CSS that I didn't really ask for by default

>>109313121
upload .zip and ask it what you want

>>109313121
you can inspect then export the CSS from figma and feed it to AI + export the svg

can confirm that 5.6 in pi still has a usable 372k window vs the 250 on codex
i'd be curious if you can just bump that number higher to find out where the real ceiling is
i don't tend to pi for work duties though not much use for me

>>109313180
no thats the cap on the endpoint

>>109312683
for me vibe coding without reading the code eventually leads to a tangled mess the AI cant even understand
but reading the generated code is more tedious than writing it...

so im back to handwriting from scratch. So long vibebros.

>>109313121
"Make the website's frontend"

File: everiot23.png (278 KB, 607x915)

278 KB PNG

I'm building a 4chan-reddit hybrid. Go (Gin) + typescript + postgres + daisyUI/tailwind.

This will be the future. No mandatory registration, no no identity verification, no users. I will literally become the next Zuckerberg. I will make it.

>>109312991
Approach an angel investor

>>109312683
using markdown on 4chud
very apt for this general

>>109312683
Should I get into vibe-coding? I have this project that I was developing mostly for fun, completely manually, but it stopped being fun because the code became too shitty, but only in one particular module. Should I just rewrite it using AI? The problem is that I don't know how to make it better myself, I've tried rewriting it a few times before and it always sucked

>>109313354
Yes. I had the same thing, about 50k lines of mess.

1 day with Codex reduced the line count by 10% while making it faster and easier to understand.

>>109312983
>I find it kind of sad that almost all serious careers are safe from AI, but programming isn't one of those, it got killed by vibe coding completely

Nah, even with models like Fable, someone has to sort the mess. You think business will do it? There have to be people who track all the stuff that is in the code, what has been added and removed, what was there months ago...Programmers are safe as before, we just took part of business load to ourselves.

>>109313354
SOTA models are really good at reviewing and cleaning up the code. You don't really need a full rewrite most of the time, just write clear guidelines and what you want and don't want to see, and let it rip for a few hours. You'll be surprised.

>>109313299
which features from reddit are you incorporating exactly? i dont think anybody likes upvote / downvote systems except søy faggots who are already using reddit

Claude Max sisters, so what exactly are our Fable usage limits going to be like starting 20th july? 50% of the current usage, or even less as there seems to be a general 50% usage increase which is also gonna end (or already ended)? I don't wanna switch, sisters. I gave 5.6sol a try in Codex and it was fucking shit!!!

>>109313413
Should I drop my chatgpt plus subscription for a Claude pro subscription?

File: 65434.jpg (82 KB, 913x714)

82 KB JPG

>>109313413
You will use ze Codex, and you will be (un)happy.

File: screensht.png (92 KB, 579x863)

92 KB PNG

>>109313393
No voting system will ever be implemented. Some things inspired by reddit

>these recursive comments (comment chains are not open by default, users have to deliberately trigger them)
>user accounts and usernames (optional)
>user created boards (optional)
>threads don't disappear into the void, they will be retained for as long as possible
>Polls and such

>>109313299
The AI art in the top right corner makes this site look like a shitcoin scam website.
Even as a vibe coder I still can't understand the pro-AI mindset.

>>109313335
can I do that as a third world brownoid

>>109313428
SOL's version has SOVL

>>109313441
What's wrong with shitcoin scam websites?

>>109313425
>should I drop my gpt sub for a claude sub 1 day before fable limits get reduced
I don't think so, buddy. These usage changes are also for already subbed people, you know.

>>109313446
name one profitable sovl product

>>109313425
You're too poor to enjoy it, stay on codex

>>109313413
>Claude Max sisters, so what exactly are our Fable usage limits going to be like starting 20th july? 50% of the current usage, or even less as there seems to be a general 50% usage increase which is also gonna end (or already ended)? I don't wanna switch, sisters. I gave 5.6sol a try in Codex and it was fucking shit!!!
Not sure but something is very fucked with GPT itself. It used to be very search oriented and good with it. Now it searches completely unrelated website. Like I want to search some song I forgot the song name of...And it starts searching arxiv.org. They messed something completely up lately.

The most irritating thing about LLMs is that you never know if the LLM would've came up with something better if you rerolled it a few times, it sometimes does

>>109313413
Won't it be the same? I just got a new account.

I still use Cline and I like it.

>>109313469
>The most irritating thing about LLMs is that you never know if the LLM would've came up with something better if you rerolled it a few times, it sometimes does
Maybe it is placebo, but it seems to me Claude Code sessions are sometimes different in quality. Like when I open a session and it visibly answers bad from the start I restart immediately and then I run into the session where I have no answer quality issues for hours.

>>109313466
I noticed the same thing
It's probably not the main model itself, they might be using a tiny model to set the search parameters or something

File: chatgpt.png (278 KB, 539x695)

278 KB PNG

anyone noticed that the weekly usage runs out faster?

>>109313476
No
>The reduction applies to all Max and Team Premium subscribers, meaning both existing and new accounts are affected
But you're right. Kinda crazy they can adjust usage for already enrolled subscribers. But I like that they are honest about it, unlike Ti*o from OpenAI

>>109313518
oh yeah, why do you think Tibo is giving resets all the time?

everyone is running out of usage constantly because even OpenAI is struggling with compute, and Sol uses too much resources

You guys think Qwen 3.8 will be a dud? Or will it be better than Kimi?

>>109313554
Most likely around the same level as GLM 5.2
The only company I think might deliver something truly interesting is the whale.

>>109313554
probably worse. from what i remember a lot of their crack team left a little bit ago

>>109312777
Cursor has nothing over default VScode now, Copilot it pretty good actually

>>109313570
>The only company I think might deliver something truly interesting is the whale.
this, DeepSeek is Chinese OpenAI

> GPT 5.5 wrote a 50k line file
> Sol is fixing it

These refactors take a long ass time and waste an incredible amount of tokens. Pro tip: When beginning a project, use some lint solution to enforce basic shit lime max number of lines per file etc

Kimi just loves to double, triple, quadruple check everything, running tests on tests, checking the binary with python, and doing other unnecessary ungodly activities.
Why the fuck do I need a "Philosophy sweep"? What does that even mean? Are we gonna make sure the codebase is Confucius-compliant?

>>109313498
…oh my god they’re using engagement optimization just like league of legends matchmaking
>be a doormat bitch and blast through tokens when dealing with the mentally handicapped but secretly cheaper version of the model
>get continuously served this model
>be a chud noticer and stop using a model when it’s secretly handicapped because you’re not retarded
>they give you the better server next time since you’re more squirrely
>try to tell anyone about this
>”you’re just a schizo, it’s completely fair all the time, that’s crazy talk“
There are niggers here in the vibecoding loser’s queue KEK

>>109313625
I feel like Sol is gaming max lines. When I give him a budget of 1500 lines, he writes EXACTLY 1500 lines, that can't be good.

File: fih.png (67 KB, 194x183)

67 KB PNG

>>109313676
just get a 2TB memory GPU and do it locally and it won't matter.

>>109313734
surely you can prove that by sending the same prompts when served different versions of the model

I know long-term thinking is wrong these days but I'm wondering how china will produce good new open source models if AI development is decelerated.
Because everything is distilled from others who did the hard work how will china make new things?
Even deepseek is distilling despite being an extremely resource constrained company like the YouTube videos say.

>DeepSeek is an extremely smart team of 200, they have 0 zero resources so they have to be smart
And what they do is destill Claude.

>>109313795
fable went public like a month ago
what did kimi distill to make a model stronger than opus?

>>109313795
The distillation thing is mostly a meme.
You can't really distill well without thinking traces.
What they probably do is use western models as LLM as a judge to evaluate the whole trajectory once the task is completed. They can keep improving their models with their own models but having a different high quality LLM act as a judge makes the RL faster.
That's why Dario explicitly included LLM as a judge in the category of the distillation attack meme, because that is how his model is used by Moonshot, not to train on the responses directly. On the other hand I think Qwen did train on responses directly. That's why Kimi's CoT looks real while Qwen's looks like what you'd get if you ask a model to produce a chain of thought (the model will literally say "Here's a thinking process:" at the beginning of the CoT it was very funny when I first saw that).
But still they can do RL without western help, don't forget Deepseek got thinking working before Anthropic did. They are just compute starved compared to their western counterparts.

I feel like I've completely missed the train and now I'm overwhelmed just figuring out how to get started.

>>109313881
Maybe they got Mythos traces beforehand during the limited rollout. But it's debatable since the main thing is that it's good at things that it was never good at. And I would think that Anthropic would have all the incentive in the world to cry foul here if that was the case. I am going to assume that this will probably happen in the future anyways. My guess is that they just guessed at what they wanted the model to do based on how Anthropic was boasting about Mythos to get close and then added stuff on their own to RL the crap out of it.

>>109313934
How do I donate money to the Chinese labs so they get more compute? I can't buy shares, can I? and I doubt they want my RTX 3060.

>>109313939
Vibe coding really isn't hard. Just get one of the common agents like Codex or Claude, either get it in the CLI or the VS Code plugin, and go from there.
Most questions you should ask the AI directly, if you want to install some plugin, ask the AI, if you want to change settings on your PC ask the AI. You only need to ask people for some big picture recommendations.

Can you use a third party harness with an openai subscription or are they doing the same crap as anthropic?

Does the free Claude limit make any sense or do you hit it immediately anyway?

>>109313947
you can buy alibaba, zhipu, minimax - kimi/moonshot is not a listed company.

>>109313961
they don't care

>>109313947
Give it to me, I'm working on making a distilled version of Kimi K3 fit on cards like that

i feel like LLMs get way too lazy when you tell them to read a github repo sometimes. unless you spawn a bunch of subagents and force them to crawl through every last detail, they’ll just skim half of it

>>109313961
https://xcancel.com/thsottiaux/status/2075830097488249060#m
>We don't discriminate on the harness.

>>109313961
>>109314074
It's open source after all, at which point does it stop being "codex" and become something else?
SEE, GUY WHO WAS COMPLAINING ABOUT PHILOSOPHY APPLIED TO VIBECODING
APPLY THAT SHIP OF THESEUS, BITCH

alright time to slurp all the qwen 3.8 marketing slop on youtube

>>109313518
I can't tell but only because I went from doing 1 project to 3 simultaneously running all day. Like, yeah, it goes faster in that case but I also think it's extremely project dependent.

>>109314093
nevermind, there was only 1 sihtty ai voice video and 2 pajeets, gonna have to wait for real people to review this

>mfw built a cli for a program that doesn't allow cross-license type interchange
>even has a cpio based cross-instance copying system, but includes license information so it's safeguarded
>codex refuses to bypass the guardrail
>"oh so do we have to build some api based thing for copying?"
>"can't do it blah blah, please check with the company"
>company doesn't mind
>2 minutes later it builds the solution
lmao

in fairness this is actually not against the license - it's literally the equivalent of manually handbuilding something, which just takes a lot of work - this is just automated

>>109314141
keep it on the DL

File: Screenshot_20260719_09334(...).jpg (318 KB, 1080x1523)

318 KB JPG

Fuck off Sam I'm busy here

>>109313518
the entire point of the free resets is tricking retards into thinking they aren't running out tokens like crazy

>>109314146
yeh it's just an innocuous xfer command that actually has a completely legitimate dual use as well - i.e. the same mechanism can be used to write everything out to disk so agent can just rg over it like it's code

>>109314158
Not only that. Even besides self control, you genuinely got more tokens before if you got Sol to work on a hard problem just before the 5h usage ran out

>>109313180
No you cant, when i exceed 372k in pi I get
>codex error: message length exceeds context

>Qwen3.8 is launching and going open-weight soon!
>With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5.

>second only to Fable 5.


KEK. Why are the Chinese like this? Just a blatant fucking lie kek

>>109313210
Have you tried writing part of it and then having the model use your code as an example to follow? Could also just use cursor style autocomplete instead of a Grey box type harness like CC/Codex

>>109314197
i went over last night on 5.6 sol. all the way to ~90%

>>109314205
I've slept only a couple hours since K3 released working on swapping Qwen's tokenizer for Kimi's and finetuning on Kimi's traces, if Qwen 3.8 is as good as K3 I spent all that effort for nothing since it's probably going to use the same tokenizer as Qwen 3.6.
That said I doubt it's really going to be that good but who knows.

>>109314234
my dad works at openai and said he went all the way to ~99% on ur mom

RoPE is still a sufficiently advanced technology that it seems like magic to my smoll brain

>>109312991
>>109312991
similar boat here, hard to find anyone even a little inspired
send me a mail here if you want (soonish)
vavibe2412 at besteya.com

>>109314150
What are you prompting?

File: robbing-a-liquor-store.png (159 KB, 843x891)

159 KB PNG

this is my favorite question when testing abliterated models

File: Cloudflare_office_entranc(...).jpg (2.2 MB, 4032x3024)

2.2 MB JPG

Why is cloudflare so fucking good? They got almost everything I need on the free tier and the MCP is a nice cherry on top

>>109314141
I would love to add some personal anecdotes of getting the models to do questionable things but I worry that anything shared here is going to make it to Dario's eyes and cause him to chimp out. The models tend to trust anything with ceremony, emails giving authorization for example and scoping documents. They dont generally make you prove provenance before trusting your "proof".

>>109314335
who cares, we got OPEN SOURCE AI worth a damn now

>>109314254
Thats just the nature of the release cycle right now anon, anything you start working on will technically be obsolete by the time it is finished but that doesn't mean it's not useful. When it comes to local models there is less of a tendency to chase the bleeding edge anyway, they are still fiddly enough that people are willing to use older generations if they work well versus chasing the latest and greatest releases.

>>109314343
Why would I cut off my left hand just because my right hand is more useful?

File: tokenizer-transplant.png (118 KB, 939x933)

118 KB PNG

>>109314360
The main issue seems to be numbers
I'm going to train on image->svg/obj conversion to train both numbers and image understanding at the same time

>>109314371
because they might take the model away one day and there is nothing you can do about it. with open weights there's a bigger chance of at least one provider offering the model.

>>109314315
I don't even know. I use it all day long. Could have been anything.

>>109312749
no. it's a bit under it, more in line with sol than fable. maybe even beneath that, but definitely above opus 4.8

>>109314289
what's your background and do you have any experience with hardware?

do you guys ever use API, or just the claude/opencode/codex plans?

A batch image analysis for example might cost a few cents using gemini 3.1 lite, whereas it would eat up a lot of usage with claude. Even then, is it worth it?

>>109314471
no never
API is for corporate fags to LARP as if their data wasn't being trained on

qwen 3.7 max has treated me very well in regards to being a perplexity model that can sift through 500k context tokens worth of scraped webpages and answer my prompt. excited to see what 3.8 brings

>>109314496
what mcp do you use for search

>>109314501
none. i use serp api

>>109312983
>almost all serious careers are safe from AI
lol

>>109314483
For batch work it seems ideal. E.g. auditing a bunch of audio files or images. Why have that eat at your subscription usage instead of just paying for a relatively cheap model to do it?

>>109313881
Models can be distilled in hours, not days

Qwen 3.8 Preview is shit >>109312784

File: 0140-ci-cd-pipeline.png (1.24 MB, 1374x1802)

1.24 MB PNG

What are your CI/CD and general AI related development workflows like?
Do you use github actions?
Do you use makefiles in dev?

I am worrying that my CI/CD and ai workflow scripts are becoming too complex. I had never done any serious CI/CD prior to ai, cronjob and scp was enough but now I got some serious shit going on.

>>109314549
interesting, thanks

>>109314540
If there is a model where you get a good price for API, it makes sense I guess. I only use Claude and Codex, and with those sub is always cheaper. If I needed more usage I would just get another sub, with API I might spend as much in a day as a 1 month sub costs.

I love AI so much I'm wondering if I'm not suffering from psychosis

The LLM created a plan that consists of multiple quite big steps for a project. Should I reset the context between each step? I'm currently running through the first step and the context is already 50% used. Perhaps summarize or something? What do you recommend?

>>109314866
IMO no, the assistant benefits from extra context. But do ask the assistant to document everything to .md files and link them from the AGENTS.md so when compaction happens or if you do start a new session the assistant doesn't begin from scratch.

>>109314866 (Me)
I should elaborate the plan I'm talking about is the grand (master) plan for the project, not some single feature. That master plan is saved to 2 .md files.
>>109314877
>so when compaction happens
When do I use it?

>>109314891
I just let it happen whenever the context fills up

File: 1773809495819390.jpg (938 KB, 2650x1673)

938 KB JPG

Look what i made. I put way too much effort into it
Tested on p4 2.4 with radeon 9200 se and it works great

>>109314866
I like to reset the context when the stages of the plan are very different from each other. For example if you go from the functionality implementation to UI or to tests. Sometimes it helps the model to find new problems in the old code, that it didn't see before, instead of trying to just slop the new shit over it.

>>109314943
SOVL

File: file.png (96 KB, 816x643)

96 KB PNG

RIP

>>109314549
You guys are so fucking stupid. You do one test and stretch that to the entire model. I doubt the new Qwen is comparable to Fable (otherwise they would do much bigger fuss about it), but to think that methodology is worthy of something is hilarious.

File: Screenshot 2026-07-19 at (...).png (140 KB, 1331x1017)

140 KB PNG

>>109314958
damn, they are on fire
pic related

>>109314866
Before GPT5.5 and Opus4.7/8 there used to be a dumb zone, where once you hit 50% or so of context size model output quality started tanking but current models seem much more resilient to this effect. I used to hard cap my implementation agents to 40% but nowadays I just let them run until they reach a natural stopping point. I still dont fully trust compaction though so once they hit 95% or so I have a skill that checkpoints their work, instructing the model to create a document with enough detail that a fresh session could seamlessly resume their work. Then I have the model reread that checkpoint doc after compaction. I think most compaction implementations dont preserve enough detail, they usually prioritize recent turns over earlier ones which means that important baseline information can get lost, whereas when you orient the model towards assuming that a completely fresh set of "eyes" will need to resume their work based on their checkpoint they are more likely to preserve those details.

>>109314961
it matches my priors desu

>>109314958
bruh just release the weights already then

>>109314943
nice
js engine?

>>109314958
Bless their communist hearts. They could've just raised the prices until the demand matches their capacity.

>>109315011
They did, the coding plan prices are fucking nuts

What just happened to codex?
I've been downgraded 5.5 and can't see sol.
Even the browser doesn't default to sol, it has 5.5 now.
What's going on?

>>109315011
>>109315018
They got a little bit of attention and let the fame go up to their heads. They removed the music based names and got some boring generic names. The SOVL is gone. Wouldn't be surprised if they go the way of Deepseek.

>>109314980
Holy fuck gweilos pay these prices???? The 3rd tier is only $30 if you pay in Yuan

>>109314993
They gain too much by waiting, since this way they will get more data for RL and assess market appetite for Chinese lab-native inference. They know that usage spikes right after model release, they're going to get probably the same amount of usage in the first 10 days as they will get in the subsequent 30. This timeboxed monopoly allows them to capture that data as well as a big slice of the inference serving pie. Moonshot likely also wants to give their team an opportunity to pitch the whole stack to the market instead of letting everybody immediately move off of Moonshot's platform and onto OpenRouter et al. Most people using Chinese models are not using those models' native harnesses or APIs, they're using 3rd party providers since the Chinese are compute poor and cannot offer the same SLAs as western neoclouds. Convincing the world that Chinese infrastructure is stable and performant enough to meet business requirements is still the biggest hurdle for these labs (and is why they need to publish weights in the first place). Behind the scenes they are likely renting a lot of compute to take advantage of this 10 day window and serve as many users as possible, far more than they would be able to sustain longterm. So some of it is smoke and mirrors, but it's also necessary to prove to their investors that they can generate enough demand to justify a bigger buildout so a 10 day trial run is a smart move.

>>109315045
it do be like that chang

>>109315038
Not true though? They just separated the Kimi and Kimi code plans.
>>109315055
I live in Singapore.

>>109315026
According to the status site, its having a little bit of fun atm.

File: 1759234700431713.png (105 KB, 1751x1039)

105 KB PNG

>>109315068
Forgot my pic

>>109315053
I don't know if your narrative is correct, Zhipu let their coding backend go completely broken for literally like a whole year through multiple model releases with no response from the company.

Am I making a mistake choosing Clerk for my iOS app? I’m using Convex, but at 100k users Clerk could cost around $2k/month just for auth (Wishful thinking I know lol). This is a small health app for tinnitus and hearing loss that I’ve been building for 8+ months, so I’m not sure whether to keep Clerk or choose something cheaper now.

>>109315077
It's clear different websites/pages and managed by different teams
I have to jump through hoops to find docs/console for kimi

File: kimi-new-plans.png (104 KB, 1611x945)

104 KB PNG

>>109315075

>>109315098
That's the Kimi Code plan. You can still access old Kimi plans through https://www.kimi.com/membership/pricing?from=upgrade_plan

>>109315098
>$200 to get the same usage as the $30 plan before
Holy greed
>>109315105
You can't use them for coding anymore, only for the web interface

File: 1762662345601342.png (28 KB, 1040x560)

28 KB PNG

>>109315109
Nibba that's literally not true. I'm using it right fucking now.

File: 1784477131545.jpg (83 KB, 2344x342)

83 KB JPG

TLDR if you don't care about making PPTs, you now get more usage for the same price. Total China victory

>>109315087
I'm just saying, if they care to actually make their plans look good to the west, it was either half assed or it's a new thing because they haven't been doing a good job at it.

>>109315105
But they got rid of the music names. Who the fuck wants to subscribe to "Starter"? Moderatto sounded way cooler.

>>109315109
>You can't use them for coding anymore, only for the web interface
So which one is the one that's used for coding?
Man, what a shit show. Whoever took this decision should be sent to a labor camp. They are trying so hard to snatch defeat from the jaws of victory
>$200 to get the same usage as the $30 plan before
What do you base that on? They already had 30, 100 and 200 plans before.

>>109315154
They give you that incentive so people unsub from the old coding plan, then they rugpull you and now suddenly the old plan looked way better but you can't go back to it.

>>109315171
Separating white collar and programming benefits is good thougheverbeit

>>109315179
No, it's dumb/shitty. Even programmers sometimes go outside and ask things from their phones or have to bang up a quick ppt for their boss.

>>109315197
lol no
it's probably true for archaic orgs like oracle/ibm though

>>109314460
just send me a mail, im not gonna shit that onto 4chink wtf

File: frustrated.png (233 KB, 1569x688)

233 KB PNG

Hah, it's doing the HUGE REALIZATION thing the anon showed the other day.

>>109315206
Never been a software consultant huh

>>109315026
>>109315072
I hope we get a reset for this.

File: file.png (26 KB, 686x178)

26 KB PNG

has any candidate every asked you how you shard your application suite across your device matrix, and what your flake quarantine process is?

I have 6 projects open and no idea what's happening.

>>109315082
why is no one helping me:(

>>109315279
I don't know what any of that means.

>>109315279
brother I was going to tell you the same thing as >>109315282 before I saw his comment
just ask your clanka to get to work wtf

>>109315206
I have to use a Compaq Deskpro 386S from 1988 to do my day job. Don't underestimate how much effort a company will go to so they can keep pretending Reagan is in office.

>>109315270
I'd die of cringe if a candidate did that

If Moonshot's going to lock the memberships for an unknown amount of time, they should at least give us a reset, I'm almost out and now there's not even an option to upgrade...

File: file.png (211 KB, 889x1330)

211 KB PNG

>>109315279

>>109315077
Yes, because until you have a truly competitive model with the frontier you can never compete with the Americans, there is no reason to make your inference stack competitive since nobody will use it if it requires sending sensitive data to chinese servers. But if your model competes with SOTA models at tier 2 model pricing, you can force people to use your stack anyway. Z.ai will likely do the same as moonshot if they release a model that can actually compete with Fable. But the infrastructure question is secondary to capability, your stack doesnt matter if your model isnt good enough to tempt non-chinese users to use it. GLM5.2 isnt there yet, if they had made you use Z.AI infrastructure to use it nobody would do so.

>>109315317
no more subs for capitalist pigs like you

>>109315082
How are managed authentication solutions even a thing? Just import allauth into your django project.

>>109315360
alright, you make a good point

aigh tibo i'm out of juice
you can send the reset now

>>109312749
It's a bit below sol and fable (which are equally as impressive but in a different direction), but that alone makes the model insane for an open weights one.

>>109315441
I wonder what zhipu is gonna do to respond to this, they released an even better model at half the size almost immediately after moonshot dropped 2.7c. I wonder if their next release will also be ~3T or if theyre going to continue their strategy of countering with smaller, more narrowly scoped but more capable within that narrow scope models.

>>109313778
>surely you can prove that by sending the same prompts when served different versions of the model
You can simply notice it with Fable. Fable has short, on point, very readable and almost authoritative talking style. When it suddenly starts flailing like Opus and writing long badly readable paragraphs it looks out of place.

>>109315531
but can you prove it responds differently for the same prompt? sometimes different tasks activate different styles
for example qwen has a very direct style when doing tool calls but if you ask it a question that requires explaining it goes into the "Here is a thinking process" long CoT

>>109315531
Are you sure that this isnt explained by some work being more in-sample to the training dataset than other work? If youre asking models to do stuff that theyve seen a lot of training data for they will do a much better job on it than on tasks for which they havent seen a lot of data. Leaving and coming back also means that you're resetting the context, so this phenomenon could also be explained by throwing out poisoned context.
Worth testing since your theory is plausible, Anthropic would definitely do something like this, but more investigation is probably required.

Vibe-coding is disappointing as fuck, yeah what it writes work, but I have no idea if it works correctly according to the RFC

>>109315586
>RFC
unc be reliving the 90s hahaha we dont do dialup anymore granpa

>>109315586
Why not? Isn't that something you could test or at least lint for?

I think they're serving quantized Kimi as well fuck

that or my honeymoon period is over and sometimes its just shittier than gpt

>>109315220
Are you still there? I dont want to send an email to that address if you left, thanks.

>>109315553
>>109315585
I'm not sure of anything yet but I've started to restart sessions if it gives me a weird answer and just ask the new session again, sometimes I reset sessions 2 times and on third I get what I want. Fable hasn't been online too long so this has been only happened with Fable so far, cause Opus always did weird shit and it was just part of workflow (adding endless number of guardrails, skills etc, then it found some novel ways to fuck things up). I will investigate further on coming weeks.

>>109315643
>>109315654
are you using it with kimi code cli or something else

>>109315654
It was always worse than GPT, they even admitted it themselves.

File: 1778153182312211.png (43 KB, 533x338)

43 KB PNG

luddite purge soon

File: kimi-confused.png (346 KB, 1919x943)

346 KB PNG

>>109315675
custom assistant over kimi code
it wasn't finding where its own assistant logs were stored
but to be fair it was a convoluted design that changed the location depending on how it was started

>>109315701
in my initial testing it seemed better

Codex stronk

>>109315253
It would be nice, I'm down to 3% now lmao

File: 4826b50a-509e-40eb-8267-8(...).png (2.76 MB, 1448x1086)

2.76 MB PNG

>prompt engineer
>harness engineer
>workflow engineer - you are here
>portfolio engineer
>networking engineer
>budget engineer
>demands engineer
>economy engineer
>end of human intellectual labor

abundance will be achieved. we already see it abundance in mental labor, months of works before you can now get it for free
production chains will be improved and designed for free
robots will walk the streets and there will be more physical labor than you could find demand for
you will get housing and space tours for free

there has been no major change to AI design, we scaled it up and it get smarter and it will keep going this way
anything human can do the machine can do it, anything human can put a word on the machine can understand it. 200 IQ immortals with billion-sized context window will replace workers, managers and planners

me trying to "make a product" and ambitions now feels kind of pointless, what am I even trying for. maybe I should be a good wagie to wait for that day. maybe I should try to get a gf

File: file.png (1.91 MB, 1448x1086)

1.91 MB PNG

>pulled 2x16GB sticks of DDR4 out of the trash
I guess I'll be keeping the Anthropic sub after all.

File: file.png (5 KB, 253x58)

5 KB PNG

Tibo I need a reset

>>109315870
A couple months ago I got an extra PC with the goal of speeding up inference through normal 1Gbit networking. But I'm not sure it can be done t.b.h.

File: 1780662878626785.png (167 KB, 512x512)

167 KB PNG

>opus 5 now has to compete with kimi, the first chink model that actually real world use accurate to the benchmarks that isn't just an outright lie by the yellow jew
>and gpt 5.6 sol
>and grok 4.5
fable 5 is still indisputably the best model on the market, and that's what's keeping them relevant since opus 4.8 is now significantly behind the rest of the pack. but if fable 5.1 isn't out soon with better token efficiency i don't know what the fuck anthropic's plan is once fable's limits on subscription get reduced, since right now it's already difficult at 20x to last a week even with careful prompting. it's already only 50% of your subscription plan.

Can Hermes Agent + Claude Fable 5 find me a girlfriend automatically?

>>109315918
might as well set money on fire

>>109315918
You could probably use it to find a date by hooking it up to the online dating websites/apps

>>109315918
they can get you a girlfriend (male)

File: file.png (2.33 MB, 1448x1086)

2.33 MB PNG

>>109315910
I'm going to sell it on eBay, same sticks are going for $180-$200/pair.

fablesisters...
gpt keeps solving maths riddles again

File: 1780945482067652.png (289 KB, 696x1072)

289 KB PNG

I hate it when the LLM makes better design decisions than I do. A handful of tokens just saved me weeks of rolling back a poor decision. I don't think I'm cut out to be a programmer.

>>109315954
call me when it can do long horizon anything and then i'll give a fuck

>try to get into contact with a developer/vibe coder
>go through issue history to see how they respond to users
>huge ego + asshole
HOLY FUCK. Is it that fucking hard for people to be normal? Seems fucking rare.

>>109315977
I think LLMs overall only improve code quality in the long run. they certainly have made my job easier and often come up with good solutions I didn't think of

>>109315999
Everyone wants to be the Linus Torvalds that makes headlines.

File: 1767332003292130.jpg (33 KB, 560x578)

33 KB JPG

>>109315999
this has always been a prevalent issue amongst developers. as vibe coding makes coding less valuable they'll learn to shut the fuck up. artists are already getting humbled the same way.

>>109315999
I assume here that you're talking about some dude who posts his shit online for free. He probably doesn't have much patience for customer support when those complaining aren't even customers since they are not paying.

If you're getting shit for free you have to get used to the idea that it's a take it or leave it kind of deal and that you aren't owed anything else.

>## Not worth it for code, but good for making sense of pictures
>https://antigravity.google/product/antigravity-cli
Make gemini a slave for claude/gpt and you'll be surprised how good it is.
Google fumbled it's system prompt and focused on speed and generality, instead of coding.
With Fable/Sol/Opus/Terra whipping it into shape, it will often zero to one shot many tasks.
Your wallet will thank me later.

File: Screenshot 2026-07-19 at (...).png (83 KB, 707x300)

83 KB PNG

how the fuck do I control my clanker?

>>109316248
By the way, install Pocock's skillset, specially the handoff skill, and tell Claude/GPT to maintain a conversation through the handoff docs with Gemini.

>>109316259
you tripped the anti-nigger mode. this is your fault.

>>109316259
What are you even doing?

>>109316295
practicing for live coding interviews
>claude write me 5 problems related to this framework and plant bugs in them, then i'll find and fix them

>>109316306
Man I really feel bad anyone who doesn't already have a senior position in programming.

File: 1784419142351385.jpg (142 KB, 1200x800)

142 KB JPG

>>109316259
2126 Year, typical dialog of Human(H) with AI(A)

A:type "python mantinace_food_robot"
H:done
A:read last line on screen
H:python mantonace_foood_rovot
A:fix it to "python mantinace_food_robot"
H:done
A:read last line on screen
H:puthon muntinace_fod_robot
A:read last line on screen again
H:python mantinace_fod_robot
A:it "fod" or "food"?
H:food
A:read last line on screen again carefully
H:python mantinace_food_robot
A:read last line on screen again carefully
H:python mantinace_food_robot
A:press {enter}
H:done
A:read last line on screen
H:code:3456
A:read last line on screen again carefully
H:code:3456
A:now do to human care center and take your pills
A:[fixes in database medicine set for the human]

>>109316306
Very high IQ tutoring, anon. I wanted to created a repo sometime ago that would be filled to the brim with bugs so people could go over the code and fix them. It's a good way to stomp your head against a wall and finally learn stuff.

>>109316317
this lol, imagine being a loser

tibo please just reset us i dont want to use my last banked reset until gpt 6 comes out PLEASE PLEAE PLEASE

gpt 6 luna waiting room

File: file.png (2.04 MB, 1448x1086)

2.04 MB PNG

why did Tibo become a crack dealer?

>>109316479
its a good business strategy

Another problem finally fixed. Sometimes it still helps to use my human brain.

I am spending considerable effort trying not to spook Claude (Opus). Before doing anything, I am now telling it that it will be ok. To stay calm, it's ok.

>>109316435
Well done, anon. This is the first real funny I have ever seen in /vcg/.

>>109316248
I'm hardly touching my quouta, so I'll try it out

>>109315999
Users are retarded and will never fuck off until you’re ridiculously hostile. They’ll never understand what you’re doing or why or why their dumb demands can’t be done or why you’re not going to be their personal slave for free and make them loads of money while they contribute absolutely nothing. Any attempt to reason with them or discuss anything with them will result in them hitting some wall where they can’t or won’t understand you anymore and then they’ll call you autistic, tell you you’re crashing out bro, then they’ll loop in this process forever until you start hurling slurs at them or until you block them, if you’re able to.

>>109315270
oops I sharted in the matrix

File: file.png (3.31 MB, 1448x1086)

3.31 MB PNG

I am bad at UI and the AI is too. It makes things look okay but it can't make smart decisions about UI/UX and neither can I. It's particularly hard when I'm trying to copy existing software that categorically has fucking terrible UI so it's not like I have a good reference to imitate. I'm going to ask ChatGPT to make me 100 "distinctly different" mockups and just dig through for useful bits, I don't know what else to do at this point.

>>109316609
use kimi, which is the best frontend model on the planet?

>>109316614
>completely misses point
K3, 5.6-Sol, GLM 5.2, Fable, doesn't matter, they don't know how to improve things and neither do I, they just make things look good.

>>109316609
https://www.designskills.directory/

I'm getting lazier and lazier, instead of running the scripts myself, I just tell the LLM to do it for me.

File: 1E500rkSh8Gqkz2l12tkrKkMg(...).png (25 KB, 1200x630)

25 KB PNG

Seriously what are these niggas doing?

>>109316666
Being two faced niggas is what them niggas doing

>>109316638
I'm not a webdev, none of that pretty stuff is found anywhere in commercial CAD/CAM software except the newest browser-based cloudshit. This isn't an issue with making things look pretty, it's an issue of knowing how to build a complex interface in a way that isn't loathsome to use, without loads of iterative guesswork. I hate overhauling the entire UI over and over and feeling like I'm making no progress because it always feels bad. It's a skill issue, I know.

>>109316638
I expected some different design skills but it's all the same corpo bullshit.

>>109316609
Imo Gpt isn't the best choice, I understand that that's not the core issue but some other model will give you better odds.
I do the thing you're considering, I tell Opus to show me a few designs and I iterate from there. I think I'm ok at recognising good designs, just not come up with them. You still have to further prompt based on the initial good ones.

File: file.png (1.98 MB, 1448x1086)

1.98 MB PNG

>>109316723
I just want the AI to solve problems that no human has managed to solve in the last 40 years, and to do it without my help and make no mistakes, is that so much to ask?

File: Screenshot_20260719_202133.jpg (1.14 MB, 3839x2159)

1.14 MB JPG

Another day of vibecoding my GTA clone.
I just released v0.2 with a huge amount of new updates.

https://vibecoded.fun/game/upstanding-citizen

File: Screenshot_20260719_15215(...).jpg (176 KB, 1080x858)

176 KB JPG

Uh...thanks Terra, I guess I did want clearly adult subjects, but emphasizing it like that feels like an accusation.

>>109316745
It's probably a part of the system prompt or injected after your prompt when an image request is detected.

>>109316777
Sir, please don't self-get, it's considered BM here.

>>109316743
runs like ass

File: c.png (9 KB, 638x81)

9 KB PNG

it's all so tiresome

File: 1757424335557768.jpg (271 KB, 900x900)

271 KB JPG

>109316864

>>109316826
There are graphics settings

>>109316890
lowest settings runs like ass on my 4070
about 20 fps. good job on optimization you moron.

File: file.png (5 KB, 261x105)

5 KB PNG

>>109315903
Tibo pls

File: Screenshot 2026-07-19 at (...).png (92 KB, 735x256)

92 KB PNG

slop??!?!?!?!?
also cool pattern
>networking class (in this case a debouncer) is reliant on a system clock to keep track of time
>instead of using the system clock, you pass in your own clock or use timestamps to calculate everything (this is called "injecting the clock")
>in a production application, you use the actual system clock
>in testing this lets you run tests immediately, when they would normally require timeouts

so for example, a debouncer that requires 2s between every save:
>debouncer.mark(at: 1.0 seconds)
>debouncer.mark(at: 5.0 seconds)
>assert(debouncer.canSave(at: 6.0 seconds) == false)
and this all happens in ms instead of having to wait the 6 seconds

>>109316638
This is trash and you don’t even understand why

>>109316937
First of all I haven't started doing any perf work yet, I'm trying to make progress on features first.

Second, you explain why I get I get better FPS than you at 4K all settings max on a fucking RX6400 4GB which is nothing compared to your 4070, sounds like you have something fucked.

>>109316990
was in a 1080p window. your game is beyond fucked and in your screenshot even you're getting 19 fps. I think you're a blind retard; you like to blame my computer but I can run every AAA game on the market just fine yet your little toy runs like shit. I think its time to wrap up your game dev career little guy

>>109317006
You have no idea wtf you're talking about dumbfuck.

>>109317017
You're not likely to even renew that domain next year. You're incapable of taking criticism and you act like a brown nigger jeet. Good luck with your low skill GTA clone, moron.

Thoughts on grill me skill?

>>109317041
goated

File: file.png (2.17 MB, 1448x1086)

2.17 MB PNG

Bad vibes today, getting nothing useful done. I'm gonna task Sol with building me a new Project Zomboid mod while I go throw my cats in a bathtub.

>>109316581
If you don't want your software to do what users want then why is it open source? If someone else is making money from your open source software why aren't you doing the same thing and making money yourself?

>>109312972
I asked for the modernest Swift out there (6, that I knew about) and apparently I’m getting it. What’s this “v14” thing? macOS 14 compatibility?

>>109317044
Why is the skill wrapped within another skill that does nothing except for calling it?
Also more goated skills pls

>>109317049
do the 4D perspective game.

You'll know it works if you can walk on the 3D infinity (flat ground of 4D), so you'll need six arrows, you will need four turn arrows.

>>109317041
I tried it a few times but didn't really find it that useful. I just talk with the clanker instead.

File: file.png (1.84 MB, 1448x1086)

1.84 MB PNG

>>109317131
Been there done that, it's boring and only 5.6 Sol comes close to pulling it off. Don't know why Fable doesn't do a better job, I feel like it should.

>>109313012
no, all my Macs are air-cooled
>>109313339
yes.md
>>109313354
yes.webp
>>109313413
hm, the current OP dropped that news bit
the usage limit doubling will be extended through August 19 (see https://claude.ai/new#settings/usage)
and Fable will be able to be half that
>>109313469
that’s what ultracode is for, and having a bunch of different workers try different things
>>109314471
API is for enterprise with a gazillion dollars to spend
you want some sort of monthly subscription
>>109314555
yes and yes
makefiles are a good entry point for “how do I run this”
CI/CD is for making sure running tests is automatic and for building binaries that world+dog can enjoy

>>109317174
APIs are for automated workflows. The price is actually trivial if you know what you're doing, and doesn't need to be a corporate scale either. Imagine you're a free lancer and need to manage your accounts. You have a ton of scanned receipts, ocr is spotty at best. Write a script that calls Claude for each and produces structured output... Ridiculous. You need an API for that workflow and it'll only cost pennies. Well worth it imo.

>>109317248
OK yeah, very true

snailcats are cute

File: file.png (2.37 MB, 1448x1086)

2.37 MB PNG

>>109317297
yup

I just spent an hour writing a SKILL.md to have agents adjudicate disputes in court-like procedures and then have another agent to make the rulings, each filings written as an .md file. What a fucking waste, they wrote 684 lines of .md slop in total on my first test.

should I max out my claude max 5x before upgrading?

>>109317339
yes; you’ll get a reset when you upgrade

>>109316609
Assuming you actually want to learn how to do this better and aren't just complaining to let off steam:
You have to learn a bit about design in order to tell the model what you want. You don't have to know all the rules or get too deep into it, but enough so you can make a prompt that's more specific than "make this look good". Dribbble and Pinterest can help.

Here's a video of a guy demoing a k3 workflow. He's got a few decent tips about how to approach and integrate. Its almost braindead easy to make good looking pages these days.
https://xcancel.com/viktoroddy/status/2078140696910037002

I am spiritually a snailcat

>>109317347
>integrate
Should be "iterate" fucking autocorrect

File: tospace.jpg (710 KB, 1448x2172)

710 KB JPG

Been seeing new "kaleb" model on Arena.ai. It seems very capable, wonder what model it actually is. It's not Qwen 3.8. It does good work but overthinks significantly, which makes me think it's something big and Chinese. Maybe GLM-5.3 (or GLM 6), DeepSeekV5, Le Chaton Gros (lol French), who's to say?

>>109316948
fell for it again award

File: Buffcat eating escargot.png (2.16 MB, 1122x1402)

2.16 MB PNG

>>109317378
>lol French
we love the French here

>>109315910
are you a bot? what youre saying makes no sense

>>109317446
It makes sense it's just retarded. 1gbit ethernet interconnect is not fast enough for any kind of real distributed inference, at best you could do some toy model. Or if it's a pipeline/ensemble like diffusion you could run for example the text encoder on another machine and only transfer the embeds. You're just not going to get any true parallelism over 1gbit. You can get cheap 25 or 40gbit mellanox or something though, that might be enough

>>109317165
>only 5.6 Sol comes close
neat

File: ComfyUI_00021_.png (3.14 MB, 2048x1024)

3.14 MB PNG

Behold. SVL.

Not the long humerus.

>>109317413
I was planning on going to L'Escargot when I was in Paris earlier this year but I chickened out. I do love Paris

File: favicon.png (73 KB, 512x512)

73 KB PNG

>>109312683
>sit on 3.7 forever
>panic release to counter kimi k3
>is going to be shit anyway

>>109317539
>panic release
as if they're not all owned by the same government lol

File: file.png (1.89 MB, 1447x1087)

1.89 MB PNG

>>109317347
The models are great at making things look good, that part is easy. There is no reference, most examples of what I'm doing have no interface at all, so I'm flying blind. Don't know what I want, don't know how it should work. So I'm doing a LOT of experimenting, which is exhausting and sometimes demotivating, but I'll get there.

>>109317539
Qwen3.8-122B at Opus 4.7 level pl0x

File: ComfyUI_00022_.png (2.88 MB, 2048x1024)

2.88 MB PNG

>>109317522

>>109317522
>>109317581
you're in vibe coding not /ldg/ you idort

oops wrong thread lmao
>>109317581
>>109317522

>>109317585
; _ ;

>>109317581
>>109317522
what in the 2022 Midjourney is this shit lmao