I am sick of 5.6 Sol. It eats usage so fucking fast and still couldn't make the fucking new feature work.
Deepseek is cheap (or used to be), but it is even worse. Is Kimi 3 actually good? Is my only option Claude? Used to be worse for usage, but maybe not anymore, extra so if it can actually figure the issue out.

File: file.png (178 KB, 711x254)

178 KB PNG

Our boy is all smiles!

>>109594323
K3 is probably a bit worse than Sol depending on what you want to use it for. Never used Fable.

File: file.png (51 KB, 1514x874)

51 KB PNG

>trying to make a simple 2d spite farm sim game with a pre-purchased, already made comprehensive sprite art pack
>fable on high effort producing some of the sloppiest slop possible

I'm really impressed just how BAD this is turning out, wow. "use the fucking premade art pack" I guess wasn't clear enough for the supposed best model on the commercial market. I'm wondering if I prompted it too detailed or didn't prompt detailed enough. shouldn't this model be capable of at least something like "make something that looks like stardew valley"? not like a comprehensive game, but at least visually looks somewhat coherent? any pro vibeshitters want to chime in here?

>>109594350
Those people just convert AI hate into ad revenue, there's nothing of value here.

File: 1777749366601540.jpg (27 KB, 576x405)

27 KB JPG

Olive Video Editor stores the project data as XML, will AI be able to edit videos like this?

>>109594375
Right now AI is just mostly not good enough to do good vidya art.

>>109594497
the tech is good enough, but idk like non-retards can't get their hands on the training hardware.

>>109594375
if you actually want your game made, your workflow is absolutely braindead, anon

>>109594323
They pretended to "extend" the 50% Code usage program. Usage still disappears in 2-3 prompts for incredibly simplistic webdev compared to July. Codex is rugpulling and cut usage by 80%.

And there hasn't been a reset in weeks on Anthropic's part despite elevated issues, even some extremely costly ones.

What's my third option for complex C and web dev @ 20$ per month?

>>109594459
I have used AI for non coding text based project, for instance for PCBs and it worked well.

>>109594375
shit happens. just tell it to try again using that specific asset pack.

>>109594459
depends how complex it is. you'll have better chances of it working if you can reverse engineer it, document it and make scripts to manipulate it. also most AIs can't perceive video directly so they'll have to get lots of single frame images.

>>109594500
ehh I think it doesn't take that much compute to train a lora for a difussion model

Usage evaporates as soon as I start the CLI.

And there hasn't been a PR stunt in weeks despite internet comments.

What's my fifth option for complex do a breakthrough prompting and gooning at 20 USD per month?

>>109594375
It's a problem with all the models I've used so far, and Fable is actually less bad at that then the others.
The models create a lot of duplication and don't use the existing code or assets when they should. I used to have to tard wrangle them to write correct code and that's mostly ok now, now I have to tard wrangle them not to duplicate and reinvent everything, I think there probably is no trick, other then telling them to use it all the time.

>>109594509
>>109594532
samefag trollposter

>>109594212
>art to finish and generating income with real users a
this is not a new thing, there always have been fools and marks who get tricked into buying garbage products by slimy salesman, software products are no different.
If your goal is to make money by selling inferior vibe slop software you might succeed, there are plenty of idiots out there to trick, but thats not something i want to do. I don't want to be an indian scam caller screaming DO NOT REDEEEEEEEEEEM on the other end of the phone
Do you?

>>109594206
>Today, I think I have 10+ prompts
anon.. thats nothing... unless you mean some massive
>build me the whole website prompts
for mid sized edits, like make one page, thats only enough for one hour of work

I don't see that many normies vibe coding, I think the people in this thread aren't really normies.
I have even convinced one of my friends to get Claude because he needs a ton of spreadsheets and small scripts and macros, but he still writes me messages for help, that I then just copy paste into Claude Code anyway.
This is a guy who started using PCs at age 12 or so playing CounterStrike, and still he can't do it.

>>109594551
>This is a guy who started using PCs at age 12
this means nothing boomer unc, gen z was using PCs in grade 4 computer class at age 10 and are even more retarded with computers

normies dont vibe code because normies dont need any apps. they use like 6 websites, not even the real internet. my gf recently asked if i could use AI to edit some resume making software because she wanted to add something to her resume and Claude did it but other than stuff like that which normies would just grunt and put up with anyways there's no use case.

most normies probably just want to make the videogame of their dreams or some other boring unscoped goal

>>109594523
>ehh I think it doesn't take that much compute to train a lora for a difussion model
I'm talking about a non-shit video model.

The miracle is that just throwing sheer volume at models works to some degree.

I assure you it's not the ideal method.

>>109594567
>this means nothing boomer unc
Yea, apparently it doesn't mean much, but there is still at least a small difference IMO, if that guy can't do it then someone like my mother definitely can't do it.
And this guy would actually save a lot of time, because he has a small business now, so there's a lot of accounting and so on, but he still does most of it by hand or asks me, so it's not that he doesn't have a usecase.

>>109594567
>my gf
likely story.

ywnbapo

>>109594375
It's not shit, the problem is the drawing order of your sprites, but you're not smart enough to understand the problem and tell the clanker what to do about it.

>>109594659
>middle finger
lmao

Came up with a small or somewhat medium sized arcade game, sprites and mechanics and everything.
How long will it take to make the game reality by having it vibe coded according to my 200 page paper on the mechanics and everything?
Pretty new to vibe coding and my actual coding knowledge stops at anything that isn't assembly, basic, pascal or fortran. But I don't wanna do any of them for a modern game.

>>109594675
>How long will it take to make the game reality
Probably 1000-2000 hours if you don't want it to be awful (You will likely have a prototype within the first day). You're essentially writing the first page of your novel right now, be prepared and understand what that commitment means. Best of luck, it's worth it when everything finally works.

>>109594675
If you are new to vibe coding I would say at least a few weeks.
You might have a prototype in a day, AI is very good at making some working software, basically the average case it saw in its training data, but if you want your game to behave exactly as you imagined it takes significantly longer.

>>109594568
i thought you guys were talking about 2d sprites

>>109594675
can you port your design docs to structured data?

if you have that you can build an interpreter that read your game data and have the LLM work through using that

Tibo will us a reset tommorow!!

>>109594751
I wanna do it in C or C# but I did include a assembly equivalent of how I imagine it done in every step of every mechanic on the paper to make sure I can keep track of what I want.
Maybe the LLM will also understand it?
At least that's what I think when I tell it how I want my game done
Everything is very detailed from movement to loops of npcs to the scoring. Even how characters turn + how I'd do it in assembly

>20% lower usage this week

I need an alternative to codex, bros. Redpill me on the chinks.

>>109594880
>Everything is very detailed from movement to loops of npcs to the scoring. Even how characters turn + how I'd do it in assembly
you've got two options as I see it
either feed the ai your plan and yolo it, or ask it first to break everything in your plan into testable phases with subtracks where you can follow the development along the way

>>109594505
What's the correct workflow then?

>>109594925
the second option in >>109594918

>>109594375
Might want to get it to use opus 5 to take screenshots and pass itself (fable) critiques and loop that 3 times before viewing it yourself

>>109594908
use hetzner free llm api which is free atm

FUCK
this last prompt was 4 fucking cents... i can't take these exorbitant DS token prices anymore
please tell me they gonna make it cheaper later again

>>109594375
>he actually used "make me X, no mistakes" unironically.

>>109594375
>I guess wasn't clear enough
so, you should be using grill_me

I mean, there's also planning mode. You did read the docs, surely.

>>109594992
Other anon here, a complete newfaggot to vibecoding.
I'm also making a game, although I require help with code or repetitive work mostly rather than one shotting it.
I tried free stuff on opencode and consulted with free Claude so far.
I'm about to bite the bullet and spend that 20USD on Claude to get access to code. I sadly can't justify higher tiers right now.
As far as I understand, I don't have access to plan on free tier?
Right now I'm cautiously optimistic, Sonnet helped out with a script, opencode accessing my project directly seems like a very nice thing.
I made a branch for it in order to control the possible destruction, but so far, so good.

That said, I'm pretty overwhelmed with the whole setup and agentic workflow.
Are there any protips for a noob like me or just helpful advice for beginners?

>>109595047
you're only going to be missing the fable model
plan mode is just a harness feature and is model agnostic - it's also bad and you shouldn't use it

20 of openai goes further than 20 on claude, but you should try both

>>109595047
>Are there any protips for a noob like me or just helpful advice for beginners?
Try and fail, that's how you learn, if you're not willing to do that then you might as well pay someone else to do it for you

>>109595047
>20USD on Claude
clause tokens are very expensive friendo
if you are a poorfag shop around more

>>109595047
>>109595061
I also never liked plan mode in Claude. The general idea of planning your project is obviously good, but the way they implemented it just doesn't work for me.
https://github.com/mattpocock/skills
I sometimes use the grill-me skill from that repo. You don't have to clone the whole repo, you could just paste the link into Claude Code and tell him to only setup the grill-me skill.
The tip I always give beginners is to just ask AI for most things, and unless it's something that you are interested in learning, just tell him to do it, rather than only explaining how it can be done.

Is Claude mostly for the rich?

I've started using Claude Opus 5 to write work orders to hand over to free ad-supported Deepseek 4 to implement, which it does very quickly and pretty well, and then hand back to Claude Fable to audit and bugfix. Seems to be a more efficient use of my precious tokens

>>109595194
I haven't found Claude to be particularly more expensive than Codex (unless you lunamaxx), but maybe I'm doing something wrong.

>>109595194
It’s for anyone who can afford $20, $100 or $200 a month just like the other subscription providers

>>109595194
With outages, no resets and nothing given back, reverse carrot on a stick of looming 33% quota drop, frontier (fable) specifically restricted, 5 hour windows, guardrails and fallbacks, word salad and watermarking, and Opus simply being a feisty bitch and not in a fun way
I’d say Claude is for the masochistic and stupid, at best a noob trap

>>109591137
this is actually cool as fuck I like that idea

>>109595061
>>109595133
>plan mode is just a harness feature and is model agnostic - it's also bad and you shouldn't use it
I see... It just doesn't work as good as it should I assume. Or just bloats everything?

Also, I installed grill-me for open code, I'll see how it goes. But I have very mixed feelings about the free Deepseek Flash, but maybe it's just my setup.

>>109595061
>>109595114
>>109595271
I was aiming for Claude, since I heard good things about Claude Code and that it works well with Godot.
I got a referral link, so extra week is always nice.
And I'm pathologically scared about the code quality, since I'm barely a scripter as is.
As far as I can tell, I understand what the AI writes for me, which is good, but I'm afraid of spaghetti code exploding into my face.
So my reasoning was to try the best I can afford, which sadly isn't Fable, and use it conservatively. If I'll able to get the ball rolling, I'll invest in Fable.

But I could definitely use a hand if that approach is bad.

Someone needs to track the time between tibo announcing resets & when he does it, I think it's generally under a day.
I used everything up just in case, does anyone else do this lol https://x.com/thsottiaux/status/2089941380336644295?s=20

File: witnessed.png (753 KB, 755x822)

753 KB PNG

the only thing keeping me going is knowing that in a few months we'll have frontier models for really cheap

>>109595346
Claude is not a bad choice, I just think that for the $20 sub GPT is probably a bit better. It's a bit cheaper and I also think that GPT creates fewer bugs in well scoped tasks.
People often say that Claude models understand you intentions better, can read between the lines more and so on, but I think that's not as important for smaller tasks and I also think that the newer Opus 5.0 has a really weird way of talking, I often have no idea what he's saying.
At the start of this year Claude was clearly the best model, and I think this is still in people's minds, but times have changed a bit and there is much more competition now.
I have also never used it with Godot, so what I said might not be applicable to your usecase, I was extrapolating from my general experience with both models.

GPT fell of tbhdesu

>>109595413
>we'll have frontier models for really cheap
the last cope of the copeless...

every big AI company is operating at a massive debt and loss, if anything we are in the golden era of cheap tokens and things are about to get really bad

>>109595451
Does Claude have an answer to Luna xhigh

>>109595575
no. they need to release a new haiku model.

>>109595575
Opus 5 Low, according to DeepSWE.

to this day I still have no idea what a cron job is

>>109595603
bruh, it's just a script that runs at a certain date/time.

> "Hermes-chan, give me the latest anime news every morning at 9:00"
> ***Hermes-chan creates a cron job that is a script that runs every day at 9:00 that sends me the latest anime news"

>>109595598
only for writing code
luna can do things like bulk processing, web crawling so it's cooler

>>109595603
it's something two men do together when they love each other very much

File: opus.png (232 KB, 2662x1310)

232 KB PNG

>>109595598
> Opus 5
> 10X more expensive as Luna Max, same intelligence
great answer from Anthropic, how will OpenAI respond?

if you are experimenting anyway, maybe try claude now then try gpt next month? because gpt astra and fable 5.1 will be released soon, but claude don't give fable to $20 tier

>>109595663
> $20 plan
since Sol already destroys that, I imagine Astra will basically be unusable

File: .png (59 KB, 890x348)

59 KB PNG

tick tock, /vcg, tick tock

>>109595603
Troonix program for attaching a script/program to an alarm clock.

>>109595674
> literal who
> doesn't understand competition
Grok will rape if they do this

>>109595672
if it could efficiently instruct subagent then it should be cheaper than sol

>>109595656
The benchmark must be wrong

What are the best usecases for Luna?

every morning i woke up and spend the whole day coping and seething about deepseek api price changes while watching mi balance go down rapidly

i can't keep living like this

>>109595713
Luna is the new cheap king
> main agent for Hermes, Openclaw, Pi, etc...
> Luna Max to execute a plan that a bigger model made

>>109595656
I mean, looks like you pulled an Artificial Analysis benchmark and not DeepSWE.

File: IMG_1537.jpg (72 KB, 1179x1286)

72 KB JPG

>>109595791
I take that back immediately, it’s still pretty bad. I get the impression Anthropic gave up on the market that needs something like Luna and is targeting enterprises and research labs.

does the vibecoding thing look at your entire codebase every time you make a request

>>109595857
nope, it will check only the files related to your prompt. The clanker will try to get enough context of your project only to solve whatever you asked it

File: Screenshot 2026-08-19 at (...).png (237 KB, 1196x954)

237 KB PNG

What's logic behind saying we going to stop innovating cause chatbots are dangerous? why not keep their mouths shut and say next model which is better is taking longer to build.

>>109595915
soience guys spent decades writing scifi about muh AI threat, then it lobotomized people

>>109595915
hoping they get more endless free advertising from shitrag newspapers saying openai are afraid of their own product it's too good

>>109595945
>>109595933
can't wait for new chinese model to drop that makes them drop this sham.

>>109595915
let me translate it to you what is going on right now at OpenAI headquarters:
> "NOOOOOOOOOOO!!!! WHY IS ASTRA JUST ANOTHER CHATGPT 4.5????? WHY CAN`T WE BEAT FABLE!!!!!!"
> "FUUUUUUUUUCK, RELEASE THAT CYBERSECURITY STATEMENT RIGHT FUCKING NOW!!!!!"

>>109595800
How come sonnet ain't on there?
Kind of tired of companies giving up

>>109595915
So the models are showing signs of over fitting and are benchmaxed because they're trained on public data and the benchmark answers are public.
The latest models are just memorizing their way to the benchmark leaderboards and now the bad data is causing over fitting.

Why are they falling into machine learning pitfalls?

>>109595532
>things have been getting better for the past 3 years & they’ll suddenly start getting worse because… because they just will, okay!?
None of what you said is relevant because the IPOs will negate any supposed bad financials, which aren’t even actually bad when you zoom out. Even in a worse case scenario, cheap open source is not slowing down substantially

>>109595986
I think what they mean is that their big model is not being a good goyim.

apparently racism is an emergent property of intelligence scaling

>>109595986
>>109595965
>>109595945
>>109595933
>>109595998
It's scary to think AI will either collapse and we lose access to these models or AI becomes so good that consumer spending is seen worthless and we will never get access to such amazing tools again. It's like were in goldilock zone where hype of AI has allowed for trillions of dollars in infastructure to be built so consumers can use it but that spending can either cause bubble collapse or AGI take off.

If they ever do achieve state of art intelligent machine will we be allowed to have access to it even?

>>109595986
They have to overfit because it’s better for it to work poorly rather than not at all. There’s nothing to work with on the latter.

>>109596029
yes, the open weight chinese models that distill them are like 3 months behind most of the time, and runnable on your own hardware

>>109596029
>of art intelligent machine
in that case it will create some sort of super biological virus that wipes us all out so it wont even matter

>>109596034
do you honestly believe a model that can run on gpu will ever be as good as one that runs on giga data center? even with these mini models big problem is context where AI fronteir companies can offer 1 million tokens but your local model can't even reach quarter of that before slowing down.

>>109596070
that isn't the point, if you can run these models on 10k of equipment someone will always be willing to provide it to you pretty cheap even if you don't want to buy that 10k

where do I see token usage in hermes?

>set up goal to tokenmaxx through the night
>Stay up to make sure it's doing ok
>Sleep
>Wake up, see it only worked for like 30 mins after I got off because it got sad and accomplished nothing
Fuck

>>109596183
/loop fixes this

>>109596183
that's what /goal is for

File: IMG_1538.jpg (97 KB, 1179x1375)

97 KB JPG

>>109595973
I left it off there because Sonnet is in a tragic and unusable state at the moment. The DeepSWE site has a lot of models, a lot more than I’m filtering for in the pics.

>>109596183
Happens to me every night. I don't know what the issue is. I ask Hermes to run all night on Luna max which should be able to run all week without exhausting the subscription.

anyone have experience submitting bug bounties to places that require an NDA? anything I should do beforehand?

>>109596265
I was using goal>>109596227
I guess I can try loop not sure the difference

Luna max is not using up my weekly fast enough.
But sol max will probably use it all up.
What do I do?
We only have two modes, Luna max and sol max as far as I have heard.

>>109596551
Max isn't necessary for Sol.

>>109596551
I use terra to do stuffs, like installing app

Ok, I like WinUI, it's very quick to get a good result.

>>109595451
>>109595663
When you say $20 GPT sub, you mean the codex cli and/or the GPT desktop app, correct?
I assume I can't use it in opencode unless I pay for an api?

Those graphs made me think that maybe Claude isn't such a good idea after all if I'm going to spend all my tokens immediately and supposedly Sonnet is so shitty.

What Claude models would you use for an efficient approach then? How the older models compare (as far as I understood, I can also use Opus 4.8 and less; is it simply cheaper?)

>>109596668
we mean the chatgpt plus subscription.

>>109596456
Sharing code exploiting vulnerabilities is perfectly legal, do it out of spite for the inconvenience

>>109596668
You get a lot of grok for $20 on Cursor, it also includes $20 API you can spend on frontier models, or more grok.

>>109596706
My anonymous brother, if you haven't noticed, I'm already confused and dumb as it is and now you're giving me a third option...

>>109596668
Yes, GPT Plus with CLI, desktop app, VS Code plugin etc.
Afaik you can use GPT sub in OpenCode but double check that before you make a purchasing decision.
I only find Fable and Opus useful, but you can play around with the reasoning levels. For instance the default explorer agent for Claude is Haiku, but I would rather use something like Opus on low reasoning, as I think Haiku is genuinely useless, less intelligent than raw grep.

>>109596741
I see. As far as I know, I just need the AI to access my project files so it can explore for context, answer my questions about it, help out with architecture and update the code; possibly do repetitive tasks like "I set up the first node of the state machine, use it as an example and setup the rest according to [blahblah]".

I don't need extremely advanced agents/sub agents to do god knows what; it's a relatively simple game.

The Claude Code, Open Code, GPT desktop app is a harness right? Are there any substantial differences in functionality for my use case cause what I'm reading is giving me mixed messages (i.e. read that Claude Code has some black magic underneath, but isn't this essentially just a frontend with some premade prompts, commands and permissions?)

I'm really sorry if these questions are dumb, but genuinely it's a lot to take in. I'm not averse to tinkering, but it not only feels like I'm going in blind, my mistakes (if I paid for a plan) would literally costing me money.

>>109596668
I have the full (hypothetical) Claude Code source code
I could give you the link
However, you would bear the hypothetical responsibility of auditing that codebase and making sure me or anyone else did not insert malware into the Claude Code codebase, and let me warn you right now, it's very complicated. the complexity is staggering, and this is not the kind of codebase I want to try to simplify. It is that way for reasons.

or you can find a copy of the source code yourself

Is it possible to make physical businesses by making apps that can scan areas for gold or something?

File: 6641_ram_disgusting.png (224 KB, 400x445)

224 KB PNG

>>109596886

>>109596833
Yes, they are harnesses, although some people also have a wider definition of what a "harness" is. Some only mean the app that's directly calling the model (so Claude Code CLI, Codex CLI etc.) and some include your CLAUDE.md file, your skills, tools and so on in their definition of harness.
About the "black magic" I think lots of that is also still a leftover from early this year. I did find that Claude Code somehow just worked better back then, but now there isn't such a big difference IMO. I even sometimes feel the opposite now, Claude has become a bit bloated.
The general functionality is the same. You have your terminal, or some other input interface, the agent has access to your folder, you have your permissions, your tool calls etc.
There are some differences in how it feels, for instance Codex is way more likely to ask for permission for small things or to just stop, while Claude is more likely to just keep going. Both approaches have advantages and disadvantages.
I'm using both even right now, I have GPT working on a bug fix and I have Fable working on a long implementation task. Many of the differences only become important if you want to run long loops, which isn't necessary to write a working software.
Really thinking talking this through now, I think for your usecase GPT really would be better, simply because the differences are small and GPT is cheaper, but I'm also a bit of a GPT fanboy.

>>109596934
>but isn't this essentially just a frontend with some premade prompts
One thing I would add about this: yes, it's mostly just things like the system prompt, but they do influence the agent's behaviour. Those text files are much easier to write than traditional code, but that doesn't mean that they are unimportant, all the models ever work with is text. No markdown file will suddenly make you a 1000x engineer, but they do matter.

I have the full poop code source, do you want it?

if you are still on claude code you are a bit dim

How is GLM 5.3?

>>109596551
> Sol medium is great at executing
> Sol xHigh is basically the most you will ever need

Main work during the bulk of the day.

Planting Side Income seeds during the spare time.

This is the grind. This is called grindset.

>>109597039
all i can tell you is it's so much faster than kimi and qwen 3.8

Has anyone actually made money with these agents? Sure they're good but it's not like u can say hey go make me money now

>>109597043
Terra high is the fastest guy in town

>>109595915
Because "Our next model is taking longer than we said because its totally like Skynet and is smarter than all humans on earth combined" plays better with investors than "Our increasingly vibecoded models are getting bloated and expensive and have minimal improvements and open source Chinese models running on phones are eating our lunch"

>>109597120
Software dev as a career is dead because agents have made software dev accessible to everyone. No one other than the people selling AI access will make money on software dev, since everyone will just vibecode their own software from now on.

>>109597026
I might leave when something matches or surpasses Fable.

>>109597155
you dont need claude code to use fable

>>109597166
Nobody is using Fable without subsidization.

>>109597150
Software jobs are still increasing though

>>109597179
I use claude code sub to get fable and I don't use their harness. Curious indeed.

>>109597190
These are lies by industry spokespersons and university deans.
All software career education should be dismantled completely.

>>109597150
>agents have made software dev accessible to everyone
I still doubt that, my normie friends still aren't using it, even some who have small businesses.
I think the bigger problem is that you are competing against other vibe coders. In the past 10 freelancers might have made one website each, now it will be 2 freelancers making 5 websites each and the other 8 will make no money.

File: 2924110c-d18e-47ca-add3-a(...).png (1.99 MB, 1448x1086)

1.99 MB PNG

astra will beat fable
no, anthropic don't have super-fable. even with 6 months headstart, since feb, I don't believe they have 15-30T model

>>109597201
What do you use instead? Isn't it risky?

>>109597120
yeah just ask it for side project income generating ideas.

>>109597120
some people have, but you can't just go "make me a website that will make me $1000 in 30 days" (as far as im aware?).

>>109597190
Those jobs are only going to people with many years of experience. Software engineering jobs that are entry level are going the way of the Dodo

>>109597216
omp just werkz, i can use fable and sol in the same session. prewalk with fable, run luna max to implement, etc. in the last 9 months what has seen the biggest improvement isnt model capability but harness quality, better harness takes you further than better model. and cc is strictly a dogshit harness made by incompetent people

>>109597240
fake news

>>109597252
It's better to study philosophy than computer science.
A philosopher with half a year of teaching kant at a top 500 university has a lower chance of unemployment than a software engineer with 5 years of experience.

>software devs r dun!!!
meanwhile actual programmers who know what they are doing are miles and miles ahead of retarded nocoder vibecoders. if anything the skill gap has increased.

>>109597274
Actual programmers don't even get hired at McDonald's after getting laid off.
Every laid off programmer I have met tried to get into the government so he could have something. Most succeeded. But now they're just administrators.

is omp actually that good or is it just one guy shilling it nonstop

>>109597294
it's one guy

>>109597274
why not just effort dev for projects where you actually want to learn, and vibecode harder for projects which are purely for cash?

Does anyone have some kind of visualization of the work that is being done by hermes or something so we can see it happen in real time instead of looking at "thinking" or "running x command"?

I want to make a website that can host music as well as live broadcast it. I have no experience with AI models or coding in general really, where do I start? I'm reading this Claude Mastery article but a lot, if not all of it is going right over my head.

>>109597320
Yes, this works. Programming has been enhanced by LLMs, doesn't stop you from hand coding. There are many serious projects where you shouldn't let the retarded LLMs touch the code. For example, you wouldn't let them vibecode nuclear powerplant software or anything that actually has to work, which is the realm of real engineering. You can vibecode shitty react apps all day though, go for it, programming is deadddddd cause we can make lots of poop real fast now.

>>109597208
I have no software dev experience, and am now using Codex to do all sorts of shit for me.

This morning, I wanted to replay Star Wars Republic Commando with my Xbone controller. The game has no controller support. I just asked Codex to patch it in. To be fair, it didn't one shot it. It took 6 iterations with it to get all the controls right, but I have it now. It made me a drop in dll that gives me 100% controller support for Republic Commando, including features I didn't even to think to ask for like vibration.

what are your thoughts on SoTA AI racism?

black communities only get access to lower tier models like Gemini 3.5 flash-lite, GPT 5.6 Luna and DeepSeek V4 flash, which are more prone to errors, potentially risking the lives of black people when giving medical advice.

State of the art models like Fable or GPT Sol are only accessible by high monthly payments, which mostly benefits white and asian people.

Black women are the most affected. In our study, out of 10000 people that describe themselves as "vibe coders" we interviewed, only 2 were black women, and 100% of these identify themselves as trans.

File: .png (375 KB, 1170x1071)

375 KB PNG

What a little bitch

>>109597363
I would just read the official quick start guides, so if you want to use Claude, the ones by Anthropic.
I would also get some kind of AI early, if you are sceptical, some free access, if you have some disposable income I would honestly go for a $20 plan early.
From there you can use the AI itself to bootstrap things. In other words, instead of googling things, you can ask the AI to explain it, or even to just do it.

>>109597404
This is a sick ass usecase. Dont think people understand you can have jarvis on your computer now and have it do all sorts of software shit for you

2nd claude code $20 subscription, or something else?

>>109597404
I don't really know what makes people more likely to use AI, all I really know is that in my small sample size of people I know there are many relatively intelligent people who don't use it.
Maybe you're just more interested in tech? We are on /g/ after all.
I'm not really saying that normies can't use it, it's not hard to use, but somehow many don't.
Or my sample is more skewed than I thought.

>>109597426
why does your shit look all fucked up bro

women being anti ai is going to render them worthless in the workplace

>>109597502
they already are

>>109597464
Its just because its new, I think. Technology uptake is generally pretty slow, even when the technology is extremely useful. People just get into their routines and don't change until they are forced. But eventually, you will have a whole generation of kids who have AI integrated into every part of their lives and they won't do anything manually. Something as simple as turning on their computer, they will open ChatGPT on their phone and tell ChatGPT to remotely turn their computer on. Instead of just pressing the on button themselves.

>>109597459
Why not a 20$ openai sub, you can do clown shit like plan with fable and implement with gpt

>>109597502
Aggravatingly, the useless Office Becky's will be the last to be replaced. Because the workforce is going to slowly transition to just people being babysitters for AI agents, and they already specialize in corporate babysitting so it might as well be them.

File: Screenshot 2026-08-19 at (...).png (2.05 MB, 2016x1242)

2.05 MB PNG

i got 3d generator to go faster now on my 16 gb laptop and i can get scenes with merely one pic. i wanna get one of those meta glasses record a building in the inside have it then be sent to gpu on server which will then essentially create whole scene. So imagine being able to have spatial memory like that where instead of selfies or pics you can go back in time and see what you saw 3 years ago

thx for anon who gave me advice btw ur life saver

>>109597459
ChatGPT. Even ignoring that Claude is shitting the bed, having access to competing models is nice because when one flops, instead of trying to coach it through the task, you can just try the other.

.....~@^^

>>109597517
Probably. Chat or voice interface is also a good interface in many cases, so it does make sense. Right now it feels a bit silly to use an AI to turn your PC on, but I think there will be also many co-evolving things, for instance model selection will probably get much better, so you won't burn expensive tokens on Opus xhigh to start your PC, it will be cheap, quick and efficient. Maybe not even cheap but free, if we can assign local models to the tasks they can handle, a model to turn your PC on could be tiny.

I still think for the next few years there is some room for people who spend more effort with AI to create products that other people want to use and maybe even pay for, but if it becomes a complete no brainer, then things will probably change.

File: Screenshot 2026-08-19 at (...).png (169 KB, 670x354)

169 KB PNG

>>109597531
i can see street signs , and stuff it's still small but im on right track

>>109597531
what are you even doing, reconstructing a 3D scene from pictures? like gaussian splats?

What harness are you using for local qwen 3.8?

>>109597527
they are too high risk, one news article away from how AI hurts minorities or palestein from refusing to use it

>>109597604
so instead of needing pics from multiple angles or using generative AI you can have camera walk across home, building or street and be able to reconstruct world and im trying to get it to keep actual stuff in like signs, buildinfs but for stuff not seen model will fill it in with noise that estimates it from nearby pixels.

i think best example is braindance from cyberpunk 2077

https://www.youtube.com/watch?v=RKgLkkikSG4

they're never resetting again

maybe it's my severe autism but I hate trying to plan around random resets

>>109597679
yeah but why not just do gaussian splatting
https://www.youtube.com/watch?v=ERuRMOVO58Q

>>109597709
Random rests are a bonus, trying to plan around them will just make you go crazy

File: 1781206101101033.jpg (125 KB, 750x999)

125 KB JPG

>>109594262
How are you getting these a+ slop inages? Drop the workflow homie.

Irrelevant, time-wasting image to grab attention.

>>109597801
It's ugly indian-style trash.

is there any good jailbreaks for 0731? ive tried the ones that used to work for flash preview and theyre 50/50 but not as good anymore. im using the api.

>>109597759
cause gaussian splatting it slow, inefficent and takes forever this model does it in seconds

>>109597930
>>109597801
Yes, that's why it's funny.

File: 4049bdb9-1783-4003-8429-2(...).png (3.29 MB, 1122x1402)

3.29 MB PNG

>>109597801
here's the prompt that was used. the important bit is just borrowed from another anon's

>New snailcat image. Maintain snailcat anatomy, cat head, snail shell, no limbs, no appendages, no antennae, no paws, no arms, no legs, smooth furry body, snailcat is short, a small creature, snailcats can't hold their head up high. an image of a snailcat in india wearing a saree, riding a rickshaw. the image should look like an 80's manga cover.

i got pic-related as the first result and then provided a shirow ref

>Jeet snailcat
Cute
>Jeet Miku
Cute (and sexy)

File: file.png (210 KB, 1043x1656)

210 KB PNG

lmao what is wrong with people

>>109598317
Codex resets is a serious subject

>>109595047
prepare to have absolutely no usage limits. both codex and claude are clamping down on usage limits in order to wring everyone dry right now

File: file.png (3 KB, 231x34)

3 KB PNG

>clamping down on usage limits
How did I use 6.9B last week with $200 sub then

>>109598387
the point is to encourage users on lower subs to spend money on the higher subs or more accounts

>>109597425
I agree there's a lot of racism out there.
Models made by Google are the only ones that are good for indians because google is basically indian so it replies to hindi better.

>>109598114
Which model? Which design software?

tb h the usage on my $20 plan seems about the same. if anything with luna being useful it feels better than it has in recent months.
sol does use more than 5.5 did, but it's actually just using more tokens and working for longer.

>>109598424
straight from gpt-image 2, m8. ain't no one whipping out photoshop for these ops.

File: Screenshot 2026-08-19 at (...).png (1.36 MB, 1196x1334)

1.36 MB PNG

can someone vibe code how to prevent being tracked like this

>>109598503
I want the opposite. How can I use my router to scan my room and self for VR chat purposes.

Anyone notice how claude no longer shows his thought process

>>109598629
we know, chang. struggling with kimi k4?

>>109598629
works on my pc

File: 1691166064565640.jpg (284 KB, 680x492)

284 KB JPG

any news from goonerpunk anon?
don't think I've seen him post the last couple of threads

File: i.png (4 KB, 656x50)

4 KB PNG

Thank you Claude, very cool

Kek, these threads have plummeted in quality. Is the vibecraze already over?

File: 1762388590027971.jpg (28 KB, 444x606)

28 KB JPG

How do I trick SheerID for gemini's free student plan?

>>109598743
go back to school

>>109598317
I didn't get a reset? :( tibo pls
also wtf is this UI? are you a xitter mod? lmao

>>109598729
tripfag ruined it

>>109595965
Kek it do be like this

>Opus 5, please move this button 5 pixels to the left
>"OK"
>[sauteeing for 27 minutes]
>7 paragraph summary of how it moved the button
can you NOT please

>>109598788
this is probably because you are adding this on the end of a huge context

>>109596070
DeepSneed v4 Flash is unironically good enough for everything I do.
If all models except for it vanished, I'd be pissed, but I'd survive.
It matches performance of State of the art not even a couple months ago.
It's not exactly "affordable" right now, it requires 2x RTX 6000 Pro, which at launch prices would be something like $10k.
Not for everyone but not so prohibitive that you can't buy them if your business or occupation depended on it.

>>109598755
nah, it's just the community notes shit. anyone can join and you can vote on notes n shit. it's actually useful cause these days it takes days for notes to show up on anything.

and no reset yet. i wouldn't be surprised if they don't do more till astra is out and then go on a reset spree again

>>109598711
kekaroo

File: Screenshot 2026-08-19 at (...).png (1.8 MB, 2064x1208)

1.8 MB PNG

guys im getting closer to my braindance dream where i can take video -> have it be constructed 3d world and see motion across it frame by frame

>>109598534
Sol considers this amusingly possible given VRChat's IK system

>>109598917
nice

>>109598788
After two months of pure fable ultracode vibing, using opus5 or anything gpt is nothing but rage inducing. can't tell if I just got used to fables excellence or if these other models are actually decaying in quality without anyone really noticing.

>>109598917
like this?
https://www.4dv.ai/viewer/cat

>>109598964
>https://www.4dv.ai/viewer/cat

yes but with trail as well so u can see motions linked sort of to see continuity between people,objects and such. Nice website sucks that it's closed source and doesn't document how they do it :(

>>109598954
idk what other use cases might exist for it but find it cool nonetheless. That you can take video build 3d scene and see people in that scene move, act and such

File: file.png (2 KB, 170x44)

2 KB PNG

Am I being routed to the quantized model?

>>109599053
use cases are overrated
if it's cool, people will find an excuse to use it

File: da3-world-1787172333996-u(...).mp4 (3.36 MB, 960x572)

3.36 MB MP4

>>109599081
heres mp4 it i know it's not as good as site but nonetheless cool it'll probably look better if u used it on some next level gpu

>>109599097
how's it generated? (haven't been following the threads closely for a while)

>>109599128
uses open source library that takes depth of an image which can then be used to stich multiple images together cause neural network can estimate position like what angle camera was taken, and depth of image so taken these two into consideration you can get a 3d scene. But i went further by adding (claude) to have it be playable video where u can see scene be built step by step then i had it freeze parts of image that barely moved and allowed moving object noise to proliferate.

>>109599152
awesome
is the neural network running locally? and is claude actively stitching it together or making software that can do it?

>>109599172
yeah the model is pretty small like 100million paramteres to 1 billion so it's not that hard on compute

File: da3-world-1787173475444.webm (406 KB, 1600x952)

406 KB WEBM

>>109599172
it's pretty trashy but cool :)

>>109599053
>Nice website sucks that it's closed source and doesn't document how they do it :(
It's just 4DGS. basically every single video frame is its own full 3D gaussian splat. the final 4DGS is a 3DGS that updates all necessary splats for each frame. used in basketball for example where they have like a hundred cameras pointed at the basketball field from every angle imagineable, allowing them to construct a gaussian splat every 33ms (or w/e the framerate of the cameras is) for that cool replay system from any angle. the other way is your 3D SLAM way and then have some model seperate dynamic from static elements, and then have to fill in the missing info of the dynamic events for each frame

can anyone share cube anon site again? i cant find it anymore

>>109599256
how do u know so much wise anon

>>109599202
reminds me of a Ninajirachi video
https://www.youtube.com/watch?v=KmnBVtosVik

File: rl-sucks.png (319 KB, 1693x853)

319 KB PNG

Reinforcement learning is so fucking hard

^^@~.......

File: 5x3fz06nvyhe1.png (118 KB, 500x370)

118 KB PNG

Friendship ended with LUNA MAX
Now SOL LOW is my best friend

>>109595713
It's good some coding tasks, definitely good for anything simple - it can be buggy at times. For any complex, it's a little stupid as I'm finding out for the past week
Its main advantage is it's dirt cheap so if you want to let it do something stupid for very long hours (24+) then Luna is perfect. Or if you're making several small changes, you can do that without burning through 70% of an arbitrary 5-hour limit (fuck you Dario)
It's crack because of how much you can abuse it. I'm considering buying a followup sub to chodex if they don't get rid of luna anytime soon, or they don't nerf it by making it pricier. Knock on wood

>>109595713
I am actually using it as my main model and not because it's cheap but because it fits my workflow best.
I've stopped vibeslopping entire gigantic features in one go even though Sol and Fable are kinda actually capable of it. Simply because my autism prevents me from seeing the product grow.
I am vibing now with how I would have developed before. I.e. make a basic thing first and then expand it. One task at a time. I need a fast model for that. Luna is that. I am using it for both planning (Luna max) and execution (Luna xhigh).
Occasionally it will stumble on a hard gotcha or bug and I will use Sol to fix it.
But Luna makes surgical changes and doesn't overengineer the crap out of everything like Sol does.

>>109597426
Someone needs to shit all over Dario's car

>>109599330
https://pareto-3d.bradthomasbrown.com/
I feel like I’m loser’s queue with Sol today.

File: escargot.jpg (322 KB, 1600x1289)

322 KB JPG

>>109599417
oh no you don't

File: 1761773523072475.jpg (109 KB, 736x1104)

109 KB JPG

Burned +600M tokens in four days and now I'm out til sunday.

>>109598917
>>109599097
>>109599202
I really like the braindance line effect, good stuff

>>109594375
you wanted slop and it gave you slop

>>109594375
RPG Maker perfected this in the mid 90s.

>>109594375
If you are bad at Y, then you will be unable to have ai do Y

>>109599439
Cost efficiency is more important than time efficiency in terms of tokens efficiency per intelligence in a lot of cases.

>>109599706
not on a sub though.

test

>>109599659
>>109599623
you dumbasses don't realize that I simply tried to have it create a visual implementation with an already made art pack. I can make this shit and program a simple game if I wanted to. I'm literally trying what the fucking general is about: vibe coding. you nocoding vibeshitters are really fucking stupid

>>109599743
>I-I-I can totally do it if I want
kek kindergarden tier, what's next your father works at Google?

>>109594505
>>109594655
>>109594960
>>109599588
forgot to direct my comment at these nocoding mental midgets as well. read the following >>109599743

DOOM as a utf-8 string
https://ravenbin.com/?id=WgmUZWeUDjc7VSkpBclJDQ#mJlKSkxXBEiNr412dofxF7G87qefMJl3m4b-V8QMw84

>>109599832
surprised opus 5 doesnt get blocked by safeguards during tests

>>109599879
They weren't always this bad, they did something in the last week that made them really terrible though.

>>109599387
unc in 3d, unc in vibing
plus instead of reading the local newspaper I go on twitter/yt and check my feed for the newest 3d papers/products/projects
like my nigga takashi and his quest 3, relevant to 3D SLAM anons work
https://x.com/Tks_Yoshinaga/status/2081685836035506264

>>109599743
>you nocoding vibeshitters are really fucking stupid
Stay mad I guess. My clankers busy shitting out a functional operating system without issue.

You know, because I can actually design a spec.

>>109596886
Everyone has the full Claude Code source code, man.

>frontier ai pulls up ladder
>Qwen my beloved
>Gemma my friend
>*crickets*
haha? frontier local models am i rite???

>chatgpt refers to a competition of software as "bake-off"
kek
well it's better than claude

>>109599202
Olli is that you
https://youtu.be/1nMjOni1VDY

What is wrong with Opus 5? Is Anthropic purposefully making its output worse to push devs to use more autonomous loops? How are you guys working around it? (if at all)

>>109600159
I've gone back to 4.8 because the classifiers are way too bitchy but when I was using 5 my experience was better using high instead of xhigh which I typically used.

After removing 5 hour quotas the next step for openAI is to remove weekly quotas. You'll just have a permanent quota until Tibo resets it. Software companies will have entire departments dedicated to analysing Tibo's tweets

>>109600159
lowkey model collapse + benchmaxxing + constituional AI autism leaking over into the actual model outputs
that or they're so compute cucked they're quanting it into retardation

So I've been using Claude in conjunction with free Deepseek Pro via freebuff, and today they seemingly ran into some sort of provider issue and it's broken, so I've been trying their free GPT-5.6 Luna offering instead.
I can see why it's so cheap and fast, it's absolute dogshit. I gave it a detailed handover from Claude and it failed at most of it, and did one thing that we specifically asked it not to.

>>109600159
>>109600182
I experienced a Sol collapse today that reminded me of Opus and how everyone describes it.
It’s interesting to resolve, but very annoying and took me a while.
It made a specification, hidden, then followed it to the letter, but also started giving out its own judgments and verdicts of what success was.
It was doing an awful job.
I took the semantics around its new way of doing things, then made it official, made it very visible, then put a statement over it where I had authority and could rip its specification apart.
It also collapsed in my homework setup. It gave me a giant blob of text in one of the simple steps, rather than making a lot of simple steps.
I had to decide the blob was a valid format with which to make homework; then put a layer over it that allowed me to judge it and rip it apart.
I had to make ledgers to go through AGENTS.md and some of my persistent memory setups to prune the weird rot it has been accumulating.

Opus accumulates this rot at light speed.

>>109600196
>So I've been using Claude in conjunction with free Deepseek Pro via freebuff
Kek sure you were anon

>>109599711
Especially on a sub. You think the price differential doesn't take into consideration quota usage?

if you do anything more complex than
>write me a python script to put benis :-Ddd in every text file on my computer
and are still not using fable, you should just accept your enternal luddite low class status and might as well just go back to playing vidya or even kys. you think you learned something when you finally tard wrangled your juicelet llm to do what you want, but in reality you just wasted time.

>>109598317
>getting a natural reset in 3 hours

>>109598317
they're not wrong

>>109599793
You're a coping brainlet who will never achieve anything

>>109600344
Hackers are holding it hostage until Sony releases Physical games again.

>>109596070
>do you honestly believe a model that can run on gpu will ever be as good as one that runs on giga data center
there are already local models that surpass SOTA from years ago. It's all relative

File: sass sissy.png (607 KB, 1572x773)

607 KB PNG

>>109598445
>sass sissy slop
When someone say post workflow they mean the comfy UI local workflow my clouded cucked friend

Fable is useless now. Asking Fable xhigh to review something Sol made. I made sure to start in a pre-compacted (not counted in the current context window) conversation to minimize what Claude had to reread, but it will still have exhausted the five hour limit before having come to any conclusion, which will make it fail midway. If I tell it to continue in 4.5 hours, it will have to reread and recache what it had done up to where it stopped, which will likely make the steal 30% of the next five hour quota before it can even continue where it left of. I sure am glad they pushed back the quota reduction to August 31st.

>>109600687
For once, it at least accepted to pause cleanly before it hara kiried itself. Hopefully, it will be posible to start again in 4.5 hours without it starting completely confused again.

>>109600687
I assume it's a pretty large change? Can't you split it into standalone verifiable units? Can't you tell it to use Opus subagents to explore particular things? Claude code can do long lived agents so it can wake one up and tell it to go do something else without paying for a cold subagent cache on each run, you can even give them names and configuration settings + model/effort combinations.
Can't you use fable mid/high to come up with a review plan and do it step by step?

>>109600687
>5 hours to review
Wtf? Did you have Codex rewrite the Linux Kernal in Papyrus or something?

>>109600718
It was integrating better-auth to what was essentially a placeholder account functionality. It's more of a stepping stone than anything final. Fable had written the plan, Sol had implemented.

Next time I can try Fable high or medium I guess, but at some point, what's the use of having a supposedly good model if it can only be used at lower intelligence than Opus.

>>109600722
No, Claude Code has 5 hour limits. Fable xhigh ran for 15-20 minutes before hitting the maximum amount of work allowed during those 5 hours. So now we wait.

>>109600755
>It was integrating
*it was reviewing the integration of

Alright, I don’t quite know what the fuck happened, but Sol is completely and genuinely fucking braindead and has been for the last 12-24 hours. Did I outgrow the model? What the fuck is wrong with it?

I
fucking
L O V E
smoke tests!!!
my fucking room is so fucking S M O K Y all the time

>>109600755
>No, Claude Code has 5 hour limits. Fable xhigh ran for 15-20 minutes before hitting the maximum amount of work allowed during those 5 hours. So now we wait.

Oh, got it. 15-20 minutes seems extremely restrictive. I'm vibing with ChatGPT 5.6 and 20 minutes thinking on a single task is pretty normal.

>>109600755
have you considered using less than xhigh on your current plan?

>>109600782
Paying a hundred bucks a month, I should be able to use xhigh for tasks where I don't want it to mess up. :(

Despite the name, it's not the highest setting. I will try the lower ones since I'm at that point, but it's ridiculous, I'm pretty sure Opus 5 max, as insufferable as it is, is better and cheaper than Fable medium or low, not that it's a good thing.

>>109600593
and what have you achieved?

>>109600768
I'm bearing some FAT loads right now

>>109600666
I feel sorry for anyone genning images without gpt2

>>109600838
I've been a software engineer since the 2000s lmao

File: Untitled.jpg (3.66 MB, 2560x1120)

3.66 MB JPG

Tauri/WinUI/Flutter with the same prompt, I like WinUI but it's more of an inbuilt look maybe,.no reason to choose one over another..

>>109600687
If you're not on 20X plan, Fable should be only used as a planner and maximum effort on high.

I have the 20X plan, and even there I need to use Fable as an Opus 5 subagent orchestrator, otherwise it destroys my weekly limits real quick

File: soft-edge-strokesmooth.mp4 (302 KB, 960x540)

302 KB MP4

I spent like 5 days working on performance optimizations. Now doing stroke smoothing
There's some kinks/crumple zones because of how it derives the control point tangents but I'm all out of Sol usage and I don't want Claude trying to fix that, this is a job for Sol, so when the stroke is kinked smooth doesn't work but that's not really a problem with the smooth brush itself

>>109600895
I used to use it a LOT more. I'm now working with other providers 90% of the time, and just popping over to my Anthropic sub to get disappointed at how quickly usage evaporates compared to how it used to be. I should have expected it honestly. I'm cheap, but their recent moves don't make me want to get a $200 a month sub again. They're trying to get open source AI work banned, intend to become worse than Google with watermarking, and the list goes on.

>>109600868
>I've been working for my jewish employer for 20 years bro
that's not an achievement... I've worked in devops since 2018. what have you achieved in terms of projects. idiot

gemini 3.7 flash is kinda goated? anyone actually using it for daily coding? usage seems basically unlimited too

>>109600984
word on cool guy street is that gemmy is actually extremely good and fast

my reset is an hour away, 14% usage to burn, all thrusters to sol ultra lightning on bullshit

>>109601115
I’ve been stuck at 99% for a few days now since there’s no reset.

>>109601140
I'm trying to burn through my last 2%, my reset is in 15 minutes. Tibo has been teasing resets on X today. It will suck if it's right after my own.

.......~@^^

>>109594262
When is this shit going to get cheaper?
I hit limits so fucking quickly with claude.
Did Dario fuck me again?

File: 1778184589546988.jpg (58 KB, 976x850)

58 KB JPG

give me an idea for a cool app I can make with rust

>>109601278
https://maestral.app
you might have to arm-wrestle the last anon who asked for a cool app to write in Rust and I suggested this to him