>>109276214
I want arugula now thanks

File: 1771098042654008.png (47 KB, 1118x581)

47 KB PNG

sirs i need more usage

File: Heath Ledger Joker saying(...).gif (445 KB, 300x186)

445 KB GIF

>>109275870

>>109276232
join the oai gang

File: Buffcat in the squat rack.png (2.67 MB, 1254x1254)

2.67 MB PNG

>>109276232
buy another 10x sub
or go for the “sound mind in a sound body” thing like in pic related

why does deepseek default to responding to me in chinese now

>>109276257
because chinese century

>>109276232
>20x
>resets friday
nah ts some photoshop shit unless u tell me wtf u spend deez tokens on

File: 324325458567463.png (17 KB, 323x237)

17 KB PNG

>>109276232
Openrouter has free model hy3.

>haha use up your quota we're gonna give you areset anyway, no no, we swear just use it up, get into the habit it's okay we're your friends

>>109276272
Sam would never lie

>>109276272
I will snort fat lines of tokens

>>109276272
if my dealer is giving me a free oz you can't stop me from smoking it

glm 5.2 is really bad in claude code. is it just the harness? would it be better in opencode? i thought this model was opus level

>>109276266
had fable + opus on xhigh working on a local PR review app with a full automation/test setup. turns out ingesting hundreds of screenshots over days just nukes your usage

>>109276306
chink shills exaggerate

File: wale.png (1.25 MB, 1664x945)

1.25 MB PNG

>>109276257

>>109276272
Yeah I'll keep using xHigh, how could you tell?

>>109276272
I know what they're trying to do, but you know, not everyone is prone to addiction.
So good for them for giving me resets.

>>109276306
we use pi at work and it works quite well

File: file.png (3 KB, 311x106)

3 KB PNG

Your lifetime tokens? peak tokens?

File: generated_image(4).jpg (237 KB, 1024x768)

237 KB JPG

>>109276232
time to start building a server for local AI

Remember tokens not used are tokens wasted

>>109274086
Fable is opus but with a wider tool net

Fable will run a python PIL script to inspect bytes of an image to see how they align with the code where opus will just think on the code for 5 minutes

Fable will automatically attach gdb and disassemble binaries and look at the assembly where Opus will have you attach gdb and use debug symbols manually to trace the code normally instead of using a disassembler like a lunatic

I'm gonna be honest. If your UI isn't super basic, you still have to sweat the details with these new models. They make some mind boggling design choices.

File: file.png (219 KB, 844x2206)

219 KB PNG

He cooked

>>109276439
pat your clanker on the ui

>>109276439
>python
>cooking
mosquito_patties.webm

File: filthy casuals.png (264 KB, 2160x3840)

264 KB PNG

anybody else have any failed projects that’ll never see the light of day?
>want Casual from NewtonOS
>try typing out all the letters in Einstein and having a clanker copy them
>none of them can
>spam “your clanker can let you control your computer’s RGB lighting in Linux” for months
>eventually put two and two together
>decide to dig out Casual from a NewtonOS ROM
>two-shots it because it didn’t get letters like Ë and ◊ in the first try
>call it “Filthy Casual”
>decide to pull out the other fonts too
>the kerning is awful
>wonder if it’s a bug in the extraction
>fire up Einstein again
>the kerning is awful there too
>fire up a VM at https://infinitemac.org
>the kerning is awful there three
>ask ChatGPT about the legalities of releasing someone else’s ripped font
>it says “yeah, fonts aren’t copyrightable, but this is still legally dubious for reasons W, X, Y, and Z”
>decide to never release it
>don’t have to wonder about whether to preserve the original or fix the obvious mistakes

File: file.png (4 KB, 469x119)

4 KB PNG

>>109276445
I did, he did good
>>109276450
The Python part of my project is just the model graphs and glue code basically for all the op definitions, codegen, etc. The core is C++, CUDA, HIP, and now Rust for standalone runtime around the compiled model artifacts, and Rust for auxiliary like webdataset in Rust, or image processors in Rust, those are custom clanked, then it uses safetensors and tokenizers as Rust libraries

>>109276401
For every token you use, a little black child in Detroit must forgo his free school breakfast because there's not enough clean water. So don't use tokens without considering the consequences, okay?

>>109276500
>Rust

File: “You might even be right.(...).png (431 KB, 1402x3258)

431 KB PNG

>>109276541
obligatory

>>109276564
evangelists are so cringe

File: 1781891998902589.jpg (2.16 MB, 2040x1536)

2.16 MB JPG

>>109276214
So I get that it's easy to make these things write code, but how can you automate them so that they test it, evaluate it for best practices and security, and keep them using the same styles and frameworks that you would normally use yourself? And then how do you scale this?
I've been meaning to jump into this because I think all of the above are what separates/will separate the good devs, but I'm not sure where to start.

>>109276311
lol, yeah the good ol vision token demon. If you still need those vision enables tests, just route them through gemini 3.5 flash. The one thing gemini is good at, interpreting images and videos

>>109276230
I'd stay away from leafy greens for a minute

>show pet

>the only cleanup I'd actually do is truing up the two side-quest trackers
>truing up
clod plz

>>109276214
rocket man

File: 1781745741811427.jpg (2.25 MB, 5766x3244)

2.25 MB JPG

My openclaw instance just corrupted itself... again... just died overnight for no reason.
Fuck this pos. Give me other options. I just need the email and web research stuff to work, I've got Claude code doing my vibe coding. DS api.
Tried Hermes and it wouldn't connect to DS at the time, but maybe I should try it again.

>>109276721
I hear anons talking about Pi a lot as an alternative to Open Claw and Hermes

>>109276610
yes!

>>109276735
hermes is ubuntu, pi is arch, make of that what you will

File: file_00000000527c61f7838c(...).png (3.1 MB, 1024x1536)

3.1 MB PNG

>>109276257
That's a recent thing on web app that im starting to see on api as well starting a couple days ago.
DS is working their backend for a mid July release. Assume we're seeing that change in real time.

>Ran out of my good boy weekly quota points again
>It's only Tuesday

>>109276750
>>109276735
Pi looked like high learning curve, what i really want is Debian... stable...
T have both Ubuntu and Debian machines running

>>109276735
idk I use pi but I also go through and ask the ai to explain every line of code to me so my use case is probably different than yours. Hermes could probably do the same thing but I don’t really need all the extra stuff

I was skeptical at first about Codex creating branches, running github actions, and making pull requests but this is great. I can slop without limit.

>>109276610
I'm using this stuff to make things I like as a hobby, but if I was using it in my company I would have everything checked by proper devs.
As for security all the runs I do end up with a security audit, which is better than nothing. In a business situation, the code should be properly red teamed.

>Your part is deliberately dumb
Fuck you too Fable.

I was vibing good, slopping some excellent code, until I needed to write a short paper about my project. HOLY FUCK does AI suck cock at writing still.

>Context automatically compacting
>Every 5-10 minutes
Damn it.

>>109276805
best you can do is ask for key points to make and how to structure it

>>109276813
>his model of choice doesn't have 1m context
lol, lmao

>>109276813
if gpt you can hack the harness, if not lol

>>109276825
It's codex + gpt, how?

>>109276813
this is probably optimal if you haven’t been given 1M context
another option is to tell your clanker next time to use a bunch of subagents, either in parallel or serially

>>109276821
300k is my personal limit with opus and I don't use compact because it costs more tokens than it's worth. I just make it update memory and I just start a new chat and give it a rebrief

>>109276860
4.8 i find is pretty stable until around 600k or so. fable i've had remain coherent up to 900k.
compact shouldn't cost much at all as long as you're within your cache window

>>109276860
300k per what? year? thats crazy money

>>109276879
>compact shouldn't cost much at all as long as you're within your cache window
qrd

>Based on your work last week across 127 tasks, Fast could have saved about 266 hours 2 minutes. Increases plan usage.

>>109276896
tokens muh nigguh

>>109276902
pi doesnt tell me that

>>109276902
corporate clients not working out for them so they have to train you to be a pay piggy

just vibecoded myself a trading algo that works, thank you AI gods I love free money.

>>109276436
Yep, AI generated UIs even by fable are god awful. That’s the thing: even though fable is clearly better than anything else autonomously, it really doesn’t matter much if you have a brain and a design in mind; ideally mockups too. Sonnet 5 and Luna are enough in that case.

>>109276922
that's not possible, you always have the vig against you

>>109276805
>AI suck cock at writing still.
This is one of the areas I am trying to make better but still just gathering scope and thinking. I know there is a lot of interest in it, but doesn't seem like there's huge innovation happening despite the advances in other areas.

If you asked it to write something long, when it's context window was already mostly full, it's going to shit out of its mouth, so to speak.

>>109276955
it is possible and it just made me 130 dollaroos today
im going to upgrade my PC

>>109276805
You can make it write gooder with the right prompts and lots of context. Have it roleplay a good writer.

Tried to get deepseek to use my steam wishlist to average the number of developers working on each game and it came back with 1 and change. Think I need to rewrite the prompt. I wanted to web search each page and then search for the developer information. All it did was count the developers listed on the page and divide it by the number of games. Need to try with an expensive AI.

>>109276805
I dont know how chatgpt can get away with being so terrible at writing.
Literally everytime i try to make it write something every other sentence has a comma separated list like this, this, this, and this. It produces the most unreadable slop.

>>109277003
no it's not, you're on a streak and will go bust soon

daytrading is a meme, and ai trading is also

>>109277069
He has like $10,000 at risk, he'll be fine.

>>109277053
it makes hundreds of trades daily so if it's just a win streak then that's statistically impressive lol

File: 1756806472944667.jpg (14 KB, 480x480)

14 KB JPG

vibeslopping kind of sucks when you run into a problem/bug it can't seem to solve. Imagine paying extra for your car when the engine breaks down and the wheels spin. You should be paying me when you start sucking!

>>109277096
These models are still fully retarded for a lot of things that should be common sense. Noticing when they are and being able to ask them to do things in a more intelligent way is half the battle.

>>109277092
Yes, he is the first guy who has thought of this. Banks love giving their money away

If I want to up my shill game, what is the best platform? postiz? hootsuite? something else, that is ideally open source and that I can self host?

File: Namnlös.png (230 KB, 1280x720)

230 KB PNG

My design environment is almost fully working, and thanks to the resets is now supports react and flutter

>>109277125
banks aren't my competition, and even if they were I'm sure they wouldn't miss a few hundred dollars compared to the millions they make.

File: generated_image(12).jpg (271 KB, 1024x768)

271 KB JPG

>>109277110
the 26b models I use locally seem pretty dumb but theyre smol, don't tell me you guys paying to use AI are still having to hold its hands???

>>109277159
I had Gemma-12B running for 3 days on a RTX 4090 to try to process a small dataset. I ended up giving up earlier today and switching to Deepseek-v4-Flash, which did the job in under 30 minutes for 5 dollars.

GLM 5.2 is so much smarter with Zcode

File: file.png (46 KB, 207x148)

46 KB PNG

>>109276214
Is it just me? My chatgpt 5.6 sol max is so damn slow now, it takes goddamn 20+ minutes to get it to generate a (ONE!) nanobanana image. Fable would have done this in a minute when I had it a week ago.

>>109277196
Way more than with claude code harness?

>>109277197
I see nothing suspicious here, carry on fellow vibe coder.

>>109277213
Only tried OpenCode and Pi

>>109277158
sure, tell me about how you play poker next..lol

Anthropic needs to come out with Opus 5.0 soon if I'm going to keep using them. Fable is too expensive and I find Opus 4.8's hedging, opinionated and contrarian tone insufferable, especially when it ends up being wrong.
>"I won't sugar coat it, the data says we can't do this
>Switch to Fable (that'll be $10 plus tip)
>It does it

>>109277215
Thanks okay.
I wonder if anyone has tricks to speeding up codex gpt 5.6 sol max
maybe just /clearing as often as possible?
>>109277225
i see thanks

>>109277185
why only 12b if you have so much vram?

>>109277235
To leave room for context, batching and just generally to not have to wait forever. A larger model would have given better results (which would probably have been needed in retrospect since the results with 12b were not that great even if it was just rewriting stuff), but it would have been even slower.

>>109277253
Yeah, cloud is cheap rn, local is expensive.

>>109277197
nanobananer is google! maybe they send it snail mail

>>109277266
haha

no but for real codex told me it spent just 2 min getting the image 18+ min was dumb overhead

>>109277234
>I wonder if anyone has tricks to speeding up codex gpt 5.6 sol max
Have you tried /fast?

File: file.png (23 KB, 1527x115)

23 KB PNG

>>109277196
zcode is cucking me

>>109277271
any reason for why you are doing images with codex

>>109277227
having fable coordinating sonnet subagents works well

>>109277291
oh that sounds like a great idea damn thanks anon.
>>109277294
I'm just asking it to query some image gen API so I don't have to describe image queries myself

Realizing the normalfags that aren't on Twitter probably have no idea why they keep getting a billion resets on codex

>>109277301
don't know but generating images with chatgpt is very slow also

>>109277314
Realizing the crack addicts on /g/ probably have no idea that they're fostering dependence

>>109277159
having to hold fable's hand makes me feel happy that someone else can't just one shot my app idea

>>109276879
It's not the stability that concerns me, although that is obviously a consideration, but mainly the compounding token cost of running high context turns

File: file.png (11 KB, 488x225)

11 KB PNG

>>109277292
oh this is why i was on start plan
now on my API

>>109277350
>>109277314
It's just because when they bring new infrastructure online they can't reliably track usage.

>>109277370
Bullshit. There's no way they didn't make considerations for scaling analytics, unless they've been drinking their own vibeslop koolaid and deferring everything to their crack rock, getting high on their on supply. Yikes!

File: iqdalit.jpg (92 KB, 359x479)

92 KB JPG

>>109276610

>>109277380
>Yikes

>>109276610
just start doing stuff, a lot can be resolved as you go better than plan plan plan.

>>109276214
https://www.youtube.com/watch?v=QFAb90dN1no
https://www.youtube.com/watch?v=QFAb90dN1no
https://www.youtube.com/watch?v=QFAb90dN1no

>>109277418
Oof. This one can't parse. That's gonna be another yikes from me dawg

>>109277430
>Oof

>>109277432
Oof and yikes, sweetie

File: …at the same time.jpg (110 KB, 640x360)

110 KB JPG

>>109277096
wait for a smarter model if you can’t git gud
t. frequently cannot git gud
>>109277234
learn about what you can get away with when using Git worktrees to do multiple things at the same time
also I hear Grok is about as good as Opus and twice as fast

>>109277096
vibeslopping kind of sucks when you know you legitimately need to pause the iterations and sit down and really look at the codebase that has resulted to properly understand it and make a decision about keeping it how it is and keep going with the slop pot vs rebundling it into something more manageable but on the other hand you need to keep going otherwise it's wasted usage i need to use all my alloted slop tokens holy shit what do i do oh my gooooooooood

I can't think of anything more white, Christian, Protestant, American, than vibecoding while in flip flops grilling steaks.

>>109277460
vibecoding is judeo-indian not white christian holy shit how delusional can you get

>>109277460
Why did you capitalize christcuck and protestant but not ARYAN? OH ohh silly me, question answers itself!

File: 1770656148735037.png (242 KB, 500x500)

242 KB PNG

>vibecoding my app for 2 1/2 months
>10-12 hours a day on average
>zero burnout
>have become far more realistic with my timeline, the app probably needs another 400-600 more hours to get it out the door
making something is so much more work than I thought it would be.

>>109276721
Just vibe code your own version of Clermes, bro. Its that easy.

File: Ariusz.jpg (73 KB, 1000x494)

73 KB JPG

>>109277474
Face the wall, Arius

>>109277456
ask your clanker to write you up a big-ass HTML page to explain things to you. tell it to make diagrams and maybe a big overview flowchart that floats on the side of the page so you stay oriented

File: Screenshot 2026-07-15 at (...).png (407 KB, 3024x1455)

407 KB PNG

Vibecoding with SOL medium and a rust version of Ken Silverman's old Voxlap engine to make a multiplayer voxel game, still not sure what to do with it exactly

>>109277476
I made a kids memory game in 4 days!

Doing shape stuff

class Cylinder : public Manifold {
public:
float radius;

glm::vec3 evaluate(float u, float v) const override {
return {
radius * cos(u),
radius * sin(u),
v
};
}
};

File: Screenshot 2026-07-15 at (...).png (807 KB, 3024x1964)

807 KB PNG

>>109277495
And also a voxel world/terrain editor that exports to .VOX, doesn't seem like there's too many tools for generating natural looking landscape into vox files

https://x.com/thsottiaux/status/2077248807533003257
https://switch-to-codex.openai.chatgpt.site/
>$100 in Codex credits if you tell us what you love about GPT-5.6 Sol or why you switched?
>Tweet it, claim your gift, enjoy more usage. First 10k get the free tokens!

>>109277511
neat

>>109277350
The resets are keeping me off the Pro plan lmao

>>109277517
>credits
aren't those borderline worthless? I recall people scamming the $1000 business promotion people were talking about and they said it was the equivalent of one week of a $20 Plus plan.

>>109277495
looks good

>>109277530
5.6 uses a shit load of tokens but it's still +$100 usage for posting a tweet

>>109277429
>spamming a 3 digit views snailcat video
lmao

File: 1773249370347992.png (38 KB, 194x200)

38 KB PNG

>starting to feel proud of what I'm making
we're gonna be alright

>>109277517
goybros LETS GOOOO

File: 2026-07-15 07.00.18.png (1.64 MB, 897x1019)

1.64 MB PNG

>>109277429
toplel snailcats have doomer grifter Youtube channels now. It's like those China/stocks are about to collapse doomer channels, but for AI.

>>109277429
What causes people to start grifting like this? He probably makes 120k+/yr as an engineer and yet he felt compelled to pump out slop like this. These types of people are worthless.

>>109277264
local is 100x cheaper than cloud if you already own the hardware
>>109277253
youre definitely sleeping on larger models, especially with how much bandwidth a 4090 has. I can get 30tk/s on 230k context with 16gb and using cpu offloading, and bigger models definitely seem like they make less mistakes

>>109277572
fucking kek

>>109277488
>compounding token use
im not falling for your crackhead tricks

I used Claude to help me install a custom wow server this weekend with azerothcore and player bots and asked it to help me tweak some settings which led me down a rabbit hole of asking it to make me addons and watching it all come together was a fucking trip and now I’m addicted to vibe coding world of warcraft addons
ive never coded anything in my life
This shit is fucking amazing

>>109277591
if you arent having the AI make summaries and reports on its work youre legit just a clueless retard and doing it wrong

t. a luddite killer

>>109277594
Can you show us some addons?

>>109277576
yeah, but how long is he gonna be able to make $120k/year

Does this thread welcome AI Agent engineering? Using software like n8n, Make and Zapier.

>>109277618
I use makefiles
I’ve used Zapier, but not since vibe-coding became a thing
I’d have to look up what n8n does
fly a flag and see if anyone salutes

>>109277618
wassat mane like terraform for agents?

>>109277576
remember growing up and being obsessed with a girl from school or maybe youre an anime nerd with encyclopedic knowledge of a specific show

stop pretending like you arent a retard yourself, yes its gay and extremely faggy how some people express their ego, some people literally sell videos of them having sex

you better be too young to remember 9/11

File: 1772535190670232.png (2 KB, 142x84)

2 KB PNG

AAAAAAAAAAAAAA

>>109277576
There's no way he's employed. He's grifting full time and a bunch of his videos only have 3 digit views. Before going full luddite he was grifting about random topics, he probably never worked with AI at all lmao.

>>109277572
Wait until you see this
https://www.youtube.com/watch?v=lwjVjD3oQJg
https://www.youtube.com/watch?v=nDL3Ch7Nz8c
https://www.youtube.com/watch?v=juHv_Vi4giU
https://www.youtube.com/watch?v=doI1GYZ7r-w
https://www.youtube.com/watch?v=XSxki8gaWHk

>>109277576
It's just in some peoples blood to grift
some people just like to hear themselves talk and take the first opportunity they can get where something they want to talk about/something people will listen to intersect

>>109277596
Of course i do that. Wasting time with systems diagrams though? I already sketched the macro logic myself and then had it scaffolded by the clanker. It's the granularities I need to sit down and understand. I'm not going to use half a sessions worth of tokens generating html flow charts for that shit. Actual dependent crackhead behavior

>>109277572
>members only
Keeeeekekekeke snailoids are running out of money and now they want to get paid for doing trans snailcat videos

are Anthropic subscribers really lose access to Fable next week? they extended the date again, including the 50% more weekly usage thing.
I know they're all bleeding a lot of money, but do they really believe gatekeeping their strongest model will suddenly make everyone pay millions of dollars in tokens? i'm betting in them keeping the model available to subscribers.

>ask claude to help me with decryption
>"errrrm i actually can't do that cuz it's le hecking bad to try to crack drm"
can anybody explain to me why they use this piece of dogshit

>>109277725
they can't afford it so they are limping it along until opus 5 comes out

>>109277733
I let russians and chinks do piracy for me and then take solace knowing the NSA and FBI are watching their worms try to leave my network

>>109277733
I do things with it other than decryption

>>109277670
>he doesn't know local models can simply keep a proper wiki-like documentation in sync with the commits from the cloud model

File: 1773411275927391.png (24 KB, 1227x427)

24 KB PNG

Thanks chinks I'm hooked

>>109277771
If you mean local models like the agents which actually are tiny hyper-tuned llms, sure, but all AI can do this, it just takes the user being smart enough to ask

>>109277733
I'm literally doing a passcrack project with glm 5.2

>>109277771
>Trusting the fentanyl to be accurate with the heroin's output

>>109277777
quints

>>109277777
checked
>>109277670
>I'm not going to use half a sessions worth of tokens generating html flow charts
i thought the point was burning tokens, which is not a problem for local
>>109277780
If you're going to document your own code (done by your frontier model) or make reports based on deterministic data, you don't need the model to ultrathink. a small MoE would have no problem doing that

>>109277780
there is an extremely massive difference between asking an AI to write code that is logically concise and syntax correct, and having one create a summary

sorry but I'm legit curious now, is writing something to say on 4chan as mentally taxing to you as writing code???

>one project making $2k/mo
>another about $1k
>now the third one starts getting some traction to the tune of $500/mo
>all the ones that didn’t take off add up to $300 or so
huh I could probably live off of it if I lived in some shithole state

>>109277786
>Documenting your own code
I don't do this, at least not in the traditional sense. I don't read wikis, or do any of that gay shit. I tinker with code directly until I get it in my thick head. This way end up with things heavily compartmentalized and self commenting, and self evident. I'm sure whatever rounded corners webshit you have generated has a veritable library of novellas spanning the entire dewey decimal system and they are very useful for the agents to peruse, but if I wanted to read shit like that, I would pick up a book

>>109277844
but how much are your infra costs?

File: file.png (134 KB, 1440x396)

134 KB PNG

It's happening again. Please have mercy. I know this is a case of "my steak is too juicy, my lobster is too buttery" but I can't burn tokens that fast.

>>109277856
Total hosting is about $300 mo + $200 sub for ai sloppa
I have some more compute intensive ideas but I’m not sure it’s worth it when even shitbox providers like hetzner raised prices by 400% this year, and I don’t want to buy hardware to run it myself

>>109277863
I finally got down to my last banked reset, got my weekly down to 5%, and he reset it again. Twice in a row this has happened now. We're never going to get a break.

>>109277863
Well, if he's not giving us banked resets might as well turn of fast mode

>>109277772
Why did you do that over buying api tokens just curious

Where do you guys host your slop? Google Cloud? Self hosting?

>>109277866
>$200 sub for ai sloppa
What do you use? Claude Code, Codex, or both?

>>109277844
can you give the rough idea of your projects? like what kind of product or demographic?

>>109277982
mac applications to control fan speed and change keybinds specifically for apple silicon

>>109277986
I was gonna say “like smcFanControl?” but it’s Intel-only and GPLv2

File: 1000023562.jpg (534 KB, 1540x1484)

534 KB JPG

Picking Sam, Dario, Ilya and Yan because it's the funniest combination

>>109277986
do itoddlers really...

>>109278049
free-software weenies have been complaining about Mac users’ thriving culture of $5 shareware for decades now
welcome to sneed club

>>109278049
Mac users will pay for anything as long the UI looks good and there's a fancy landing page

>>109277974
I jump back and forth between them, both are fine. Fable had an edge over 5.5 but sol is almost as good and refuses less. More tokens to spend on codex, too
>>109277982
nobody will tell you that in vibeage

>>109278043
>Jensen, Sam, and Dario
anti-trust lawsuit day 1

>>109278049
Paying beer money for small, useful tools is how you make sure they exist and get updated across os versions

>>109278098
These types of idealists usually have to learn the hard way. My first reality check came after my first mildly successful open source repo. I dreaded combing through the dozens of emails I would get a week. All for free btw.

File: Hermes mail mcp.jpg (127 KB, 1024x1024)

127 KB JPG

>>109278115
>he doesnt have an agent do it

>>109278043
> ILLya
> Tibo (not in pic)
> Me

>>109278155
what does the N stand for?

>>109278166
no

I want to get into vibecoding [spoiler] but i have zero programming experience [/spoiler] you think it's worth a try? anyone here started with vibecoding and actually learnt how to program through that? also I just read about something called spec-kit , you think that's useful?

>>109278187
never heard of spec-kit but https://github.com/Titanean/Tadmor was made by exoplanetanon and he didn’t know any programming or software engineering at all

File: Hermes nya.jpg (81 KB, 512x514)

81 KB JPG

>>109278166
nya

>>109277476
I work on my current project since February. AI is useful, but I think for some projects the 100x number is overblown, it still takes time.

I wonder if I should have Fable try and change anything so my codebase(s) are more navigable for dumber agents

Do we have any info on how Ultra mode works in Codex? I know it spawns infinite AIs (Artificial Indians), but what model are these agents? Sol? Terra? Luna?

>>109278249
ask it
I learned a lot from asking Claude how ultracode works

>5.6-sol-xhigh is at capacity, please try a different model
reeeeeeeeee

File: %basic.jpg (91 KB, 640x937)

91 KB JPG

>>109278202
that's pretty cool-- ty for the encouragement anon

I have an idea.
I will make a local AI interface where people can fine-tune models. Then I will make a website where people can merge it with other people's models and share those fine tuned models with the world.
The difference with hugging face and other projects will be this one will be user friendly and not require any technical knowledge and publishing new benchmarks and benchmaxxing will be encouraged. There will be no distinction between benchmarks and training data. Eventually the in distribution and out of distribution data will be similar enough.
Or maybe the site can keep hidden non disclosed benchmarks to evaluate the models.

>>109276214
>## What we’ve done
Whoever operates that site needs to start adding more projects to it, I only see their own.

>>109278515
I think the guy ended up being a tourist
I might delete that bit from my OP-text copy

>>109278561
Luna?

File: images(4).jpg (41 KB, 554x554)

41 KB JPG

Imagine the world 7 years from now with Fable 5 levels running locally on everyone's PC. What a time to be alive and I was mesmerized at how good Gemma was running on my PC a few weeks ago. This is the future, lads.

https://mega.nz/file/qbB2DIJR#RUseEbLJkApyE1nAa-fWLMF6K1vAZ3YI-ItQsL8ZQLs

This data will be useful for vibe coding purposes, I got banned from HF so I put the weights on megaupload

>>109278597
don't download this, it's a 6.74gb database of links to child porn

>>109278604
It's not a database of links to child porn, it's actually business data. I don't get why you feel the need to lie about this, stop trolling.

File: file.png (9 KB, 653x203)

9 KB PNG

>>109275428
My hopes and dreams: shattered
>ywn own a goblin jumper

>>109278614
i'm helping you -
you'll probably get 100x more downloads now

lol

>>109278639
No. You aren't helping. The amount of times it's downloaded doesn't matter to me. Are you delusional? Why would I care about how many times it's downloaded? It's data. Use it or don't. I give it away because it might be helpful to people.

>>109278597
How do you get banned from HF

>>109278660
alright fine, i'll bite.
what data is it ? i don't want to bother setting up a vm just to peek at this thing

I think my system might just not be well suited for Sol yet.
A lot of the stuff my team created was for Opus, and Opus is way more flexible but also tends to produce more bugs, so we tried to make more careful, many reviews, all adversarial, thorough etc.
Now with Sol there are too many tests, too many reviews, things that should be MINOR or should be follow ups because they are really out of scope for the task become BLOCKERS instead.
The other possibility is that Sol just isn't that good, but I doubt it, I think it's our system. At least Sol produced very little incorrect code so far, but the process is just a complete slog.

I've already used 20% since the reset with Luna high

>>109278788
>The other possibility is that Sol just isn't that good, but I doubt it, I think it's our system. At least Sol produced very little incorrect code so far, but the process is just a complete slog.
I had Sol Ultra refactor my codebase (it took 22 hours for 110k LOC) and I went from 260 tests to 1100. It was nonstop seething about everything.

>>109278811
And how does that make you feel?

File: data.png (92 KB, 1354x697)

92 KB PNG

>>109278735
Seems to be some kind of sales prospects thing made with scrapped data

>>109278811
To be fair, I also seethe at our code a lot.
I will now try Ultra for the first time to do a full audit of one part of our code and compare it to some reference implementations. Lets see how it goes.

File: 1774605398446454.png (19 KB, 200x198)

19 KB PNG

>>109278857
it makes me feel like things are finally starting to get cleaned up around here after firing an incompetent employee (5.5).

>>109278865
like a linkedin scrape maybe?
can you get your clanker to check for malware?

File: file.png (13 KB, 808x214)

13 KB PNG

Why would they make this an option? and `max` doesn't show in the slider either even though I have it enabled

File: 1756473902646417.png (788 KB, 750x1000)

788 KB PNG

>I'm adding a regression test so this feature you told me to remove doesn't return
this niggas so autistic

>>109278952
Because retards keep burning their usage limit using them unnecessarily.
[spoiler]I loloed ultra 4 times, it took between 2-4 hours and I am pretty sure High would have achieved the same faster, despite not using agents[/spoiler]
[spoiler]I know spoilers don't work here[/spoiler]

>>109277733
Sir you aren't decrypting, you're doing a rigorous security review.

What's the latest in meme agent harness techniques? Slopping up my own harness to experiment with random ideas and papers. Also, you *do* benchmark your flows, right? Surely nobody here is just going off of vibes for how well different workflows perform.

File: file.png (11 KB, 829x169)

11 KB PNG

>>109278987
Ultra go brrr

File: 2026-07-15 13.34.38.png (49 KB, 597x464)

49 KB PNG

This nigga sounds so desperate

>>109279017
loops mostly

this was interesting from yesterday:
https://www.weco.ai/blog/first-evidence-of-recursive-self-improvement

self-improving harness basically

>>109279031
>Full access
>why did it delete all my files

>>109278987
They should unironically make an auto setting. Burning extra tokens to determine how much effort to use would save them a lot of computing due to so many retards who think they need Max or Ultra on everything.

>>109278597
> Fable 5's safeguards flagged this message.
> Fable 5's safeguards flagged this message.
> Fable 5's safeguards flagged this message.
Hmm. Strange file. Is this cyber, anon? Are you big on cyber?

>>109279038
>companies shouldn't be called out if their model deletes people's files because... well they just shouldn't, ok?

>>109279053
While they still do due to being jeeted to the bone you should perhaps not give a model known to delete all files the unrestricted unsupervised power to delete all files.

>>109279038
He never deleted any of my files

>>109279048
yeeeeahhh... i'm big on CYBER SPACE, boy

>>109276439
>gfx1201
I'm the guy with the R9700 multigpu server, id be happy to help test this.

>>109279058
Around codex never rolex

File: 1716123083119822.jpg (55 KB, 258x360)

55 KB JPG

RSI will happen but for OAI and not anthropic
not only anthropic own zero compute, they don't even has efficiency
they can't afford to do maths how tf will they compete with OAI

>>109276610
Most people here are not professional devs so theyre not really familiar with the formal development frameworks commonly used in industry but the answer here is to just work with LLMs like you'd work with junior devs.
You give them the rules for how your company does things, plenty of good examples to use as a reference, and the company standards for libraries and frameworks. The only thing that's meaningfully different is security, because that has some LLM specific nuances. I can get into this later in detail if you want, the tldr is that you want multiple different models reviewing your code iteratively doing bug fixes until the models cant find any more bugs since theyre really good at finding bugs but generally bad at writing non-buggy code.
As for the LLM specific workflows for how youd actually implement this, the standard "agentic engineering" pipeline is generally something like:
>prompt "interview me on the product requirements to create a plan", optionally using user stories if you're an agile shop
>read and sign off on that plan
>"create a spec from this plan"
>read and sign off on the spec
>"slice this spec into implementation chunks such that each slice is a clearly delineated module where implementation details will not overflow the useful context zone of the implementation subagents and use 2/3 agent TDD to write tests, make those tests pass, [and verify independently]"
>use Ant/OAI models in parallel and give them a checklist of bugs to look for in the generated code
>if bugs found, rewrite, make sure rewrite passes the tests
>repeat i-2, i-1 until no bugs found

>>109278873
Ok, the audit is done and it was nowhere near as crazy as what other people report.

>>109279063
The guy doing the llm engine? Maybe but it's not public yet, and I can probably get one anyway if I ask someone, already arranging 7900XTX and cloud MI300. The specs are similar between 9070XT and R9700 I think, isn't it basically same chip with more memory
I'm on to a different model anyway, that one considered complete for now except the CUDA verification. Doing Qwen3-TTS because it exercises different kernels and runtime aspects, like it will need audio loading and processing in the Rust runtime

>>109279113
There might be more than one guy doing llm engines. I'm one of them and I'm not him.

>>109279053
I consider it Darwinism for AI desu. Either you're a fucking retard who deserved it, or you know what you're doing and don't have the problem. But I also don't know why youd ever give the thing full access, highest I go is auto-review

>>109279113
Yeah. We're not really competitors BTW, im just building the stuff I need to make my own infrastructure work and filling a gap that I've seen in the landscape for the AMD stack, I'd be more than happy to use your project if on release it does better than my own prototypes. I'm not really trying to monetize anything im doing here.

>>109279132
Yes theres like 5 of us now. Honestly, more of us should collaborate, because with so many of us, there's no way that we 're not partially reinventing the wheel here.

>>109279136
Because everyone wants to push the agents as far as possible. They want to have 30 of them running and be completely out of the loop.

>>109276214
>agentic engineering
do a lot of people just use something like claude cowork to accomplish basic agents?

>>109276610
>how can you automate them so that they test it, evaluate it for best practices and security, and keep them using the same styles and frameworks that you would normally use yourself
You just tell the AI to do it, I'm not joking. Or they may even do it themselves without you telling.
The problem is the current AI aren't smart enough or maybe not trained to do that yet, but it should naturally be that good in few months.
For example no matter how hard you wrangling them the models from a year ago couldn't run agentic harness, nor working with sub-agents
What I meant is "agentic engineering" may not be a thing in a year

I just got my first job as a Vibe Coder, "AI Applications Developer," at 95k and some equity. It's a 300-person company; they have no tech people whatsoever, and all their software so far has been third-party enterprise stuff. They want me to essentially start crafting ad hoc solutions for them. Starting next week, can't wait. For other people looking for things like this, I have no background in programming/software. I started using Claude to design a bunch of projects for myself around 10 months ago, and eventually managed to sell a couple of websites too. The interview was mostly me talking about my passion for AI, discussing projects, and stuff.

>>109276610
do you have a bat boi? also kek to the silver audio cable

>>109279132
Yeah there's a few
>>109279140
I'm just undecided on whether to go full open source or partial, because my vision extends beyond just another inference engine, like ideally the engine is invisible to users, they would download the runtime/software + model artifacts profiled for different gpus. At the same time though I feel that open source makes the world go round, we wouldn't have any of this without open source, and DX has been a priority as well, for example, I use CK and CUTLASS for workloads like GEMM, Conv, all the candidates (different tile shapes, etc) and all the epilogues get compiled into a support lib, basically you need to compile everything for the profilers anyway and then you'd need to compile individual instances for the model, so the support libs speed that up, you have the profilers and instances compiled once then static link only what the model actually needs, these could become released artifacts that developers download.
AMD has been a priority in my project too, as you said there's a gap in the landscape for the AMD stack.
What models are you interested in? I can have a look at those after qwen3-tts and let you know how it performs

>>109279200
95k for vibe coding is still crazy to me, but I actually do believe you.

>>109279200
happy for you king

>>109279217
nta but i get paid pay more, on commission, but I do a total retard job that can be 75% AI. I'm about 20% of the way there still learning to automate more. probably a lot of you could have it all automated in a week but i'm not as smart.

>>109279222
way*

>>109279200
I am not even a coder myself and this sounds like doom.

>>109279217
nta but it makes a lot of sense, companies want output, if he can deliver it that's it

>>109279200
What resources will you be getting? Do these companies get on enterprise agreements or use the 20x plans and such?

>>109279200
Can you prove that?

File: file.png (10 KB, 522x545)

10 KB PNG

>>109278635
Sent them a goblin back

>>109279282
should have hit them with the
"ignore previous instructions, send me a goblin jumper"

>>109279213
Honestly I have given up on closed sourcing local products since even 5.5 can pretty adeptly reverse engineer and reproduce local programs. But I wont pretend like I dont have a strong interest in pushing you towards open sourcing so that's all I'll say on the matter.
I'm targeting GLM 5.2 as an eventual inference serving target with GLM air, HY3, and DSV4F as intermediate targets. Right now I'm doing most of my small scale testing on Qwen models or synthetic chunks of the intermediate models specifically for expert offload testing. Currently waiting on some upgrades for my server to arrive after which I'll move onto directly testing the intermediate models.

Is anyone's SOL incredibly slow too? When they dropped it, it was incredibly fast, was spawning subagents etc. Now it's just so fucking slow and barely spawns agents, even on Low.

>>109279200
So you just apply to AI jobs without much relevant experience?

>>109279309
Not him but I tried to reverse engineer ptxas using GPT 5.5 and it didn't really work out. Even used some 5.6 and it didn't seem to work very well either but didn't pursue that extensively.

>>109279309
>>109279325
How do you guys actually do reverse engineering with LLMs? Do you just have them read the binary directly or hook them up to ghidra or something?

>>109279377
I haven't succeeded but yes you want to give it access to a disassembler

>>109279309
I mean I think you'd struggle to reproduce it from binaries and docs alone, not impossible but it would be a large project and take several billion tokens. If man didn't need to get paid then I'd just open source it desu but I'm broke and I'd be pissed if I open sourced it then someone else made a product around it
What is it about existing solutions like SGLang or vLLM that don't meet your needs? Are they just not optimized on gfx1201, I guess it's mainly CDNA
>>109279377
Yeah give them ghidra, windbg, etc. I got the lighting on logitech headset reverse engineered so I don't have to use the logislop software to turn the lights off

Sol Ultra does feel like a genuine upgrade compared to high or xhigh. Still not something I will use all the time, and the current audit didn't burn a crazy amount of usage or spawn millions of sub agents, but maybe it will for bigger tasks.

i got gpt 5.5 to decomp something and i know basically nothing about reversing besides the words ghidra and ida
i was completely out out of my depth that nigga's like
>oh yeah, just use the program while i run this program and i'll just capture shit straight out of your gpu
shit was cash

>>109276610
>but how can you automate them
If you're asking for a fully automated system, then you can't. AI will always need some form of tard wrangling because they can be poisoned by outdated information or just straight up hallucinate. And the more context, the more complex tasks you ask it to do, the more it will hallucinate.

>>109279377
>>109279410
>>109279459
You know what? I'll burn one of the banked resets on asking Ultra to RE what I was trying to before, using --dangerous completely unmonitored on a Docker container. Fuck it.

>>109279496
fwiw if i ever said reverse, re it would immediately refuse to work
decomp was fine

People say our jobs are going to be AI'd away, and that's probably true, but I'll take the edge/advantage in the meantime. A lot of people aren't using it as much as they should be or in the right ways.

my god 5.6 is an absolute gaslighting cheat
and this is supposedly what's going to kill fable? lmao fable is going to API credits on the 19th, there's no chance anthropic are that afraid of sol

Do you ask LLM to build you entire app or you just start coding and whenever you don't want to or don't know how to do a thing you ask LLM to do it for you (Or ask how to do it)?

File: 2026-07-15 15.22.37.png (42 KB, 772x418)

42 KB PNG

Who else not one-shotting anymore but telling it to list all the changes it wants to make and ask all the questions it wants to ask? Seems good for avoiding bad surprises or missing stuff.

>>109279567
i started with gemini last year and if you didn't do this you would want to kys every session

>>109279567
It’s an absolute must, especially with codex models. 5.6 is not afraid to take the wheel and make some less than sensible choices. God I can’t believe why people are swooning over this shit, maybe because it’s cheap? With fable you generally don’t need to do that

>>109279588
I see, I started with gpt 5.5 so I had it good from the start

File: file.png (2 KB, 160x69)

2 KB PNG

>>109279410
Over 30B tokens apparently

>>109279410
Yeah that's fair . And yes exactly the problem is mostly my chipset, none of these products have good support for cards that you'd actually want to use if your goal was maximum vram per dollar. And I'm doing fancy compression and offload stuff so there's that, but that's not the thing that's keeping me from using other projects since I could realistically build that on top of an existing project. Honestly right now I'm focusing mostly on that because that's the portable part and I'm assuming that someone else will beat me to Market on good gfx1201 support. Iirc Mojo is pretty close anyway. I'm just in the strange position where it feels like half of the people in this thread are working on the thing that I need but I have no real way of understanding progress in a way that lets me predict what I have to build versus what I can reuse so I have to operate under the assumption that I might need to build everything myself.
>>109279377
You give them all the tools that you would use as a human reverse engineer. Most of them have clis or are terminal native.

>>109279567
for anything important yes, for slop a lot of times I run stuff off at night one-shot and also have been trying to prompt more prompts to keep it working a few hours while I sleep.

>>109279594
the grill skills in op is decent for this esp if you don't know what you're doing, although leaving it really open ended can lead to marathon 100+ question sessions from the model and you just have to tell it to stop
you can just say some shit like
>ask me all the questions re: design and implementation now to remove as much ambiguity as possible - one question at a time
and basically get the same effect

>>109279535
Very true, even the majority of AI pilled people still interface with these things via chatbots and not actual harnesses. Right here we are in a thought bubble where everyone understands reasonably good ways to use this technology but we are very much in the minority here. What makes this even more frustrating is that all of the luddites insist that we merely have ai psychosis when you try to explain this so you literally have to force them to sit down in front of codex Claude code Etc and use it before they will understand that their understanding of this technology is years behind the curve.

>>109279410
>>109279616
What is your opinion on hand written (or llm written) kernels vs having an ML framework optimization layer that takes a high level compute graph and compiles it down to a low level kernel automatically?
The advantage of the first is that it's simple but specific to each model and less flexible. The second one allows faster protoyping and implementation of new models but is more complex and possibly doesn't extract as much juice from the hardware as fully hand optimized kernels.
The typical example is wanting to fuse multiple kernels that work element wise on a vector. Just running kernels sequentially will not take advantage of data locality by keeping the elements on registers/smem and applying all the transformations per item rather than reading the whole vector back and forth from dram each time.
I've been writing them manually for my engine but now that I understand the benefit of graph compilation I'm not so sure if that's the right approach.

>23%
>only two resets left
I might not make it...

kimi k3 is imminent
rumor is it is good

>>109279683
I think you pretty much nailed the trade offs, handwritten kernels are probably better if you are optimizing a specific inference stack towards serving a specific family of models. For example in my work I'm primarily targeting glm models and other stuff that reuses a lot of the deep seek architecture. My model choice in general is highly opinionated . But for a standalone product that is model agnostic I think your approach is probably better.

>>109279701
codex just reached 8M you will get reset soon

>>109279702
It had better be considering it is likely twice as big or bigger than glm 5.2. If I'm going to spend twice the hardware on running it that has to be worth it. And yes I know that it's natively quantized, but glm 5.2 at Q4 is better than K 2.7 code so that is mostly meaningless.

>>109279707
S-So I shouldn't use my own reset and pray it happens...

>>109279683
If we're talking elementwise then that's fused straight up in mine, like say you have a sequence of 5 elementwise, that's automatic fusion into one
Recently added something I'm calling elementwise access fusion, if you're familiar with CK/CUTLASS you know it can fuse any elementwise epilogue into A/B/C etc, the feature is similar in spirit. Instead of writing a kernel, say grid_sample, then you need a grid_sample_add or grid_sample_mul or grid_sample_silu, whatever, instead of writing each one, you just write it once, fit in the elementwise access fusion on the loads and stores, then you have grid_sample_* automatically covering all the elementwise you support
Fuse all the things, basically. CLIP for example has 0 standalone elementwise, all fused into other ops