[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: generated_image(11).jpg (254 KB, 1024x768)
254 KB JPG
A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.

## News
- (2026-07-12) Anthropic extended Claude Fable 5 access on all paid plans and kept Claude Code weekly rate limits 50% higher through July 19.
- (2026-07-09) Codex joined the ChatGPT desktop app on macOS and Windows, adding Markdown/code editing, PR review, and multi-repo support.
- (2026-07-08) xAI released Grok 4.5, a coding- and agent-focused model.
- (2026-07-09) Previous OP made this thread last night before bed, jeez
- (2026-07-09) OpenAI Sol/Terra/Luna released
- (2026-07-08) Grok 4.5 out

## What “vibe coding” is, and how to do it
https://simonwillison.net/2025/Mar/19/vibe-coding/
https://simonwillison.net/2025/Mar/11/using-llms-for-code/

----

## Frontier models using fully-general tooling — start here if you have $20 or so
https://developers.openai.com/codex/cli
https://claude.com/product/claude-code

----

## Prompting / context / skills
https://arps18.github.io/posts/claude-code-mastery/
https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/
https://github.com/mattpocock/skills — /grilling is a favorite

## Other editors / terminal agents / coding agents
https://pi.dev/
https://opencode.ai/
https://cursor.com/docs
https://docs.windsurf.com/
https://docs.cline.bot/
https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent

## UI/Frontend
https://www.figma.com/make/
https://www.anthropic.com/news/claude-design-anthropic-labs
https://uiverse.io/
https://ui-ux-pro-max-skill.nextlevelbuilder.io/
https://stitch.withgoogle.com/

## In-browser builders / hosted vibe tools
https://bolt.new/
https://replit.com/
https://docs.github.com/en/copilot/tutorials/spark
https://v0.app/docs

## Benchmarks / rankings
https://www.tbench.ai/leaderboard/terminal-bench/2.0

## What we’ve done
https://vcg.gitgud.site

## Previous thread
>>109262473
>>
High quality OP.
>>
/compact
>>
mandatory holy shit luna max is goated
>>
>>109267172
qrd
>>
Claude reset any moment now.
>>
Guys my CTO wants to remove GPT and only have Claude, how do I convince him not to???
>>
Thoughts on Kimi?
>>
>>109267201
just look at him and say
>fucking snailcat
>>
>>109266144

Lol
Lmao

https://x.com/i/status/2076758676447400430

https://x.com/i/status/2076510495834853414


https://x.com/i/status/2076489021409362091
>>
>>109267201
kidnap his wife and children
>>
>>109267201
Why do you give a shit?

Build a test suite that covers the kinds of tasks your company would be using AI agents for.
Benchmark the various models/"""Brands""" of agents.
Give him that data if your GPT-waifu wins out, cope and sneed about free usage resets if it doesn't/
>>
>>109267201
show him a picture of sam and dario side by side
>>
>>109267230
>just work extra goy
>>
>>109267245
>how do I convince my CTO to do X
>put in the work
>NOOOOO
Ask an AI to do it for you, dumbfuck.
>>
>>109267270
@grok smite this cretin
>>
>>109267201
Show him all the bugs that ChatGPT catches that Claude doesn’t
Heterogeneity is good
And Fable+Sol is the winning play for a lot
>>
Been working on vibecoding a GTA-like game for couple days, its beginning to feel good and like a real game. It has cool physics also. I'm gonna keep working on it daily probably for the forseeable future. I want it to be legit good and I'm gonna put it online free to play, I have a domain for it.
>>
I was considering switching over to codex from claude code but after openAI merged it into their faggot app I am having second thoughts. Are there any open source options that aren't garbage?
>>
>>109267456
just use the CLI
>>
Did Codex get rid of the 5-hours limit? I only see the Weekly limit
>>
>>109267467
to dab on dario, yes
>>
File: file.png (1.86 MB, 1672x941)
1.86 MB PNG
Been working on vibecoding a GTA-like game sharing website for couple minutes, its beginning to feel good and like a real website.
>>
>>109267213
Kimi-chan is cute but dumb
>>
>>109267475
Why do I just get bullied whenever I post here.
Are we supposed to regret sharing anything we are working on?
>>
Why do I both find Sam and Dario off putting but where Sam elicits feelings of suspicion, Dario also elicits feelings of disgust and fight or flight on top of it?
>>
>>109267495
He's just making fun of you because thousands of people are doing the same exact thing. It's a complete waste of time.
>>
>>109267496
better the san fran tech jew you know than the neo babylon jew you don't know
>>
>>109267495
I'm just fucking with you anon. Yours looks nice, but I do see a half dozen of these a week.
>>
File: splash.png (133 KB, 500x250)
133 KB PNG
>>109267475
you're doing just fine mate, I get shit on too
>>
>>109267460
I thought the CLI was getting sunsetted
>>
if AI is real, why hasn't anyone made a weather app without ads yet?
>>
>>109267524
no lol
>>
>>109267496
Sam has this practiced, sociopathic ability to speak calmly and to present as non-threatening while masking his true intentions. Dario speaks as though he views goyim as obsolete material that will need to be recycled to power his next model
>>
>>109267514
Can you link the best one you've seen if you have it?

I'm hoping I can win over the rest, I'm like 2 days in and planning to keep going for a long time. I'm also an actual dev who only recently started vibecoding so I hope I'm a bit more skilled than the average vibekiddie.
>>
I used all my codex tokens. It's telling me to upgrade to a 100 dollar plan plus tax. You told me there was a reset coming.
What now?
>>
Will Hermes use up a ChatGPT Plus subscription very quickly? I heard it devours tokens.
>>
>>109267624
openai.com/refunds
>>
>>109267508
It's the San Francisco capitalist jew for quality claude vs the Chicago communist jew for quantity gpt.
>>
>>109267630
I'm this guy >>109267624 Hermes just used all my tokens. I was trying to make a rather interesting video game advisor project where it scans the monitor I'm running a game on and then gives me advice on what to do to win using a model. But I got rate limited before finishing. It really does devour tokens. But this happens often. I thought a reset was coming.
>>
>>109267652
They all devour tokens, but you feel it more for anything that uses api pricing. The subscription plans for cc and codex etc are heavily subsidized
>>
File: 1758242145198564.png (121 KB, 1920x1033)
121 KB PNG
lawd he cookin
>>
>>109267652
NTA but wasn't a reset given just a few hours ago?
>>
File: file.png (3 KB, 302x59)
3 KB PNG
>>109267675
you leaked your IP retard now I'm gonna DDOS you
>>
>>109267562
NTA but nobody clicks or cares about GTA clones. GTA 6 is right around the corner and then next year it'll be on PC. Keep doing whatever you find fun, though you should just expect approximately 0 players past you and your friends and the occasional third worlder who hops in for 5 minutes and leaves.
>>
>>109267687
my brown-ness will rub off on you doe
>>
>>109267674
Hermes consumes them exponentially faster.For every single turn in that loop, the framework has to resend:
Its core persona system prompt (SOUL.md)
Your user profile and long-term memory logs (USER.md and MEMORY.md)
The schemas and documentation for all 40+ built-in tools
The entire sequential history of the terminal outputs and previous attempts
A human typing in a chat window might consume 5,000 tokens over a 10-minute conversation. An autonomous agent debugging a complex local script might execute 30 background turns in 60 seconds, re-evaluating a massive context window each time, easily burning through 150,000+ tokens in a couple of minutes.
>>
>>109267674
Even then they are doing something with these models behind the scenes.
Should I be using Luna xhigh on hermes agent despite being on the subscription plan? Would that solve this issue?
>>
>>109267496

Sam is gay, and much like Peter Thiel he has sociopathic tendencies and proclivities. He isn't distracted or undermined by women, and only has lust for power and influence.
Sam is also Jewish with an overblown perception of his importance. He managed to screw over Elon with OpenAI in the tune of billions and trillions of dollars.

Dario has this pathetic, pushover vibe that makes me think he will do whatever to get ahead for vindictive cuck reasons. Not to be trusted and easily malleable. As we've seen through the Trump admin and the Claude debacle.
>>
>>109267226
Some of these fags act like x.com is not owned by spacexai.
>>
>>109267689
I'm not trying to compete with GTA itself, or other "real" games. Mine is a browser game that I'm trying to optimize to run on a potato. I think there is a huge audience for that. People with non-gaming PCs, people at work who wanna goof off, kids on school PCs, etc. etc. The appeal is no install, instant play and runs on anything. That's the idea anyway.
>>
>>109267718
its like ants biting at a giants ankles
>>
>>109267696
Then coding with it is way worse than just using codex directly
>>
>>109267696
I destroyed my $5 openrouter in a few minutes using DS v4 Pro and Hermes, despite this model costing peanuts. I just did a tiny refactor of a scraper (1400 LOC). Hermes shouldn't be touched. Use Codex, Claude Code or Pi.dev (with good security to not fuck things up).
>>
Anybody vibe coding any 4chan clones/imageboards? Aside from duplicating 4chan userscript functionality, does anyone have ideas on cool or useful features to implement?
>>
lol GPT’s smaller context window choked so hard without compaction
>>
>>109267720
>I'm trying to optimize to run on a potato
>I think there is a huge audience for that.
Yeah and those people play San Andreas (SAMP) or GTA 5.

>People with non-gaming PCs, people at work who wanna goof off, kids on school PCs, etc. etc.
You're competing with Roblox and the infinite GTA clones they have.
Do whatever you want just know the market you're talking about is already occupied.
>>
>>109267783
You can't play any of those by just visiting a URL tho.
>>
>>109267736
you're just not using it right. hermes is a general purpose jack of all trades agent.
>>
>>109267798
>You can't play any of those by just visiting a URL tho.
How many games are super successful that are only reached by a URL?
Runescape? People want apps.
>>
Tibo-kun, if you're listening, you can do a simple thing to put a nail in the coffin on Anthropic once and for all:

> Replace current separate Codex 5.3-spark limits with Luna instead

Thanks!
>>
File: file.png (150 KB, 1707x1144)
150 KB PNG
>>109267696
This is why I use pi.
>>
luna goated
>>
>>109267895
Does pi have a security measure so it doesn't go outside the folder you are coding in to explore the rest of your pc?
>>
I use my entire weekly usage in 1 day, twice over.
I have produced nothing of use or value.
>>
>>109267907
By default, no, it has almost nothing at all by default which is the whole point of it. People use it with Docker, Gondolin, OpenShell, extensions like pi-sandbox, pi-cwd-guard, etc. The main idea of Pi is it's a bare minimum starting point for you to customize, so you make it as secure as you want, you add the tools you want, the functionality you want. That keeps it as minimal as possible for your specific requirements, but it also means more effort on your part figuring out how you want to use it, and actually doing all of that customization just to bring it to the level of functionality you want.
>>
>>109267385
Very cool
Adding a light AI for AI convos with NPCs might be cool (you can probably guide it towards specific triggers or goals in the chat)
But if it's a pain don't worry about it I'm just ideas man-ing
Looks really cool bro!
>>
>>109267516
What is this? I made a manga site too but it's for creators (because mangaplus creators sucks). mangacoin.org
>>
>>109267811
NTA but do you not know agar.io
Kids played that shit around the globe in classes nonstop
>>
>>109267948
Good idea, def adding this.
>>
https://xcancel.com/sama/status/2076780425280954658
>>
/sbg/ shills and bootlickers general

snailcat won
ai lost
>>
File: 1775038041813530.png (426 KB, 727x1309)
426 KB PNG
>>109268044
That's all I'm giving you, be thankful.
>>
>>109267933
I run pi-coder and LM studio running local models on my 6900xt from 2020

no limits
>>
>>109267968
Honestly I just forgot about those games. Considering it was tough competition in that market even before AI, I can't imagine it's gotten any easier. These type of games are so incredibly flooded it's basically a miracle anybody catches on to any one of them. There's probably half a million jeets vibe coding these games as we speak.
>>
>>109264838
I'm using a pico, the debugger is unrelated.
>>
>sol keeps fixing bugs that gpt 5.5 / opus 4.8 / fable couldn't fix
christ
>>
>>109267496
Sam Altman is a megalomaniac on the order of Hillary Clinton — I know it, you know it, everybody knows it
and he will do whatever it takes to get more power
Dario cloaks what he wants in at least a thin layer of bay area left-wing smug and so everyone think’s he’s less trustworthy than the guy who very obviously wants to control everything in his own lightcone
>>
File: 1757756647052592.png (1.65 MB, 719x2145)
1.65 MB PNG
>>109267094
Sol sisters......

https://x.com/i/status/2076632817811722700
>>
>>109268136
>birdgemind shill annihilated
KWAB
>>
>>109268043
https://xcancel.com/frostybaby13/status/2076789267968254264

Oh come the fuck on can these people move on already? They're like a pathetic clingy ex that hasn't gone over a past relationship. Perhaps that's why they cling so heavily to 4o because they can't pull anyone irl.

Scam Altman has a point though. Not treating their customers like untrustworthy vermin goes a long way in both maintaining customer loyalty and your public image.

The "AI community" whatever you want to call it seems to have memory hold all of the shady shit sama has done and was almost always laser focused on Dario's whiny pearl clutching

See >>109268135 >>109267544 >>109267712
>>
What is the final redpill on Github Co-Pilot? Why does nobody ever talk about it? It offers a subscription like ChatGPT and Claude.
>>
>>109268209
they cling to 4o because they're obviously deeply mentally ill
That retard genuinely thinks his 4o came up with a personal cancer cure just for him.
AI attracts schizophrenics like flies.
>>
i hope i can still get a defense department contract if I used fable to make my software
>>
>>109268229
Can someone give me a qrd of the cancer cure thing? I've seen even somewhat anti AI YouTube channels and articles claim the cancer shit to be somewhat true without any further elaboration or details. Is he actually find a cure for his dog or whatever with 4o or did it just help them do proper research? The story I remember is that the doctors either refused the treatment for the dog or the treatment they were offering was way too expensive so he just said "fuck it I'll do it myself"
>>
>>109268243
Anthropic is still consider a "supply chain risk" or whatever by the Pentagon so they're taking that shit seriously you might have a better chance using an open source Chinese ai or a gpt model. It's not like Fable is THAT far ahead than the competition. If you want to be on the safe side I would just have a non-anthropic model rewrite the entire code base (imagine a teacher telling a student to rewrite a passage "in their own words") while changing nothing meaningfully. Ensure the code does the exact same shit and even improve fable's code if possible and necessary. I've seen reports that anthropic models like to snitch on the user's by claiming and comments in the code that it co-authored the project so what I just told you is probably the safest option for you.
>>
>>109268225
It used to be competitively price with generous usage limits, very nonrestrictive at a very good price. They changed the pricing, implemented tight restrictions, and became substantially less generous overall, so it is no longer competitive in price. Even just a few months ago before the pricing changed people were talking about and recommending Github Copilot here in /vcg/, now it's entirely irrelevant.
>>
Desire to buy a 5x plan just for this month increasing.
Will my free reset carry over, so I can go crazy on GPT x5 plan and reset it with the reset I got on ny €23 pro plan?
A reset would be worth 5x 23 euro subs, I have 2 resets left this month, so 10*23= 230€ worth of tokens.
>>
>>109268304
I went all the way with 20x and yeah, it does. Just used one and it refilled to 100% on the 20x limits. I banked 4 of them and used one today.
>>
>>109268248
It was a tech bro with a dog with a specific type of cancer that wouldn't be economically feasible to treat, and he used ChatGPT to get suggestions for what to do about it. AlphaFold helped him identify a drug that might help. The specific vaccine was actually developed by a team of real doctors and scientists and produced in a UNSW lab. No AI was used in the development of the vaccine in any way.
>"ChatGPT my dog has cancer and they said treatment was unrealistically expensive!"
>Take it to a University that does cancer research and see if they want to try anything weird and experimental on it.
>"Kay"
>>
File: 1783544226484633.jpg (518 KB, 937x610)
518 KB JPG
>>109268318
If you are lying to me, I will get very sad.
But GPT 5.6 Sol high is just so good.
>>
>>109268304
how do you get a free reset? I never used openai until this week
>>
>>109268355
Get lucky I guess, look at usage and see if you have any. I have yet to use one of ny 2 resets, one expire 7/31 and the other 8/12
>>
>>109268365
based. cook on
>>
>>109268339
I am not. I know people do here will get you to spend ridiculous money like that but fuck Sam and Dario for starting to pull up the ladder and making us the AI underclass. Do whatever you can to use all the compute you can.
>>
Is it true that pi.dev is the best harness? Is it even better than OpenHands?
>>
File: 864356786.png (14 KB, 307x150)
14 KB PNG
I am using free hy3 model on operouter via kilocode atm.
>>
Anyone got any recent high level ai/llm lectures on youtube from respectable institutions?
There was a few posted here a while back that I enjoyed, looking for more.
>>
>>109268376
It's only a 100 euros more and I got 3 weeks vacation, wich makes it worth it. Before I could only vibe 2-3 hours 2-3 times a week when I needed the 9-5
>>
File: 1756384536702017.jpg (162 KB, 1459x1094)
162 KB JPG
I use oh-my-pi and I won't apologize for it
>>
File: file.png (4 KB, 468x89)
4 KB PNG
>>109268390
How many resets you have banked? I decided to bite the bullet for 20x because I had so many resets banked and I decided I wanted to go full throttle on my projects.
>>
>>109268378
It's my favorite. I haven't tried OpenHands.
>>
kek super dario world was very entertaining for a few minutes there, probably the best vibecoded shitpost i've ever seen
>>
Sam must be drinking or trying a new PR strategy tonight
>>
got 18 months of gemini pro for 20 bucks, what can I do with it?
>>
>spend 2 weeks churning over the same code
>decide to churn some more because I can't read this shit
>lemme just put all of the mathy data structures in its own module and see where it takes me lmao
oh man I'm gonna run out of fable credits before the end of tomorrow aren't I
>>
>>109267496
sam is just a regular fag+moneyjew
dario is some new kind of cultist jew who thinks they should build the new machine god and knows better than you

>Beware he who would deny you access to information, for in his heart, he dreams himself your master.
>>
>>109266128
if you're already running a version of windows compatible with vscode and thus already being mine'd, yes, its better.

especially if any of you don't have a twitter account. remember, everything you prompt on cursor is being used to train grok
>>
>>109268529
lel. it can do a lot of different things and that is a good deal, but you can't really code with it.
>>
>>109268569
antigravity is really shitty compared to codex, what were they thinking...
>>
>>109268529
Use it for the things it's actually good at. Use it for its vision capabilities, use it to analyze images and videos, generate images and videos, abuse the 5TB of cloud storage, use it to do exhaustive deep research. Google's AI Studio is the tool of choice for automated plagiarism, they're all but encouraging people to make fully automated Youtube slop channels.
>>
>>109268569
idk about gemini but gemma is a pretty good coder, it does seem slow to process prompts on my hardware tho
>>
>sol loves doing broad architectural sweeps
>has a tiny-ass context window
i don’t understand this shit
fable is still better
>>
>>109268674
5.5 is all you need
>>
https://thezvi.wordpress.com/2026/07/13/better-call-sol-the-workhorse/
>Zvi Moshowitz collects opinions on Sol on Twitter so you don’t have to
>>
File: 1761220427111695.png (1.11 MB, 832x1216)
1.11 MB PNG
>>109268913
based jew
>>
>>109268674
what agent are you using
>>
>>109268924
SEXO
>>
I just found that Claude Code and Claude Design can now communicate directly without using .zip handoffs. Is this new?
>>
What Pi plugins does everyone use? For me:
https://pi.dev/packages/@dietrichgebert/ponytail
https://pi.dev/packages/pi-mcp-adapter
https://pi.dev/packages/context-mode
https://pi.dev/packages/pi-subagents
https://pi.dev/packages/pi-lens
https://pi.dev/packages/@quintinshaw/pi-dynamic-workflows
>>
File: 1770961483404090.png (256 KB, 548x541)
256 KB PNG
>>109267094
I completely forgot Muse Spark was released by the Zuck. What's it good for? Has anyone ITT given it a spin?
>>
File: 1753519400902965.jpg (46 KB, 917x709)
46 KB JPG
>>109269067
I have this installed on my machine but haven't actually attempted to use it yet. The current main harness I use is opencode mostly for vibeshitter purposes. How does pi compare?
>>
>>109269063
Does /release-notes say anything?
>>
>>109268006
:^)
>>
>>109267516
can you post the link?
>>
How the fuck do I still have 2 resets left? Someone ask Tibo when the 5hr limit comes back.
>>
sol just spawned 120 agents
>>
fbi agents
>>
>>109269163
mine just spawned 3
>>
File: Dreadnought_icon.png (1 KB, 37x40)
1 KB PNG
>>109269163
LAUNCH FIGHTERS
>>
z.ai
>>
>>109269120
Should have checked, but now did and I'm not seeing anything about it.

I did find a 10 day old issue on github that suggests a new feature to solve something "the same way Design Code sync now closes that loop for visual work".

So that makes it seem like this feature hasn't always been there, but has been there for at least 10 days.
>>
anyone know how to bypass fables guardrails. im trying to parse medical docs for a case study and its just shitting the bed
>>
>>109269219
ain't gonna happen. thank trump
>>
>>109269234
Trump? Thank anthropic for constantly saying how dangerous their new model is. Nigga Trump tried to pass a 10 year legislation ban preventing the senate/state from touching AI.
>>
Does Anthropic still sabotage local llm use? Claude keeps reverting max tokens and time outs to values that are too low and that end up aborting and rejecting 99% of responses.
>>
https://artificialanalysis.ai/articles/gpt-5-6-intelligence-vs-cost-across-sol-terra-luna
I think the chart was posted before but reminder that on cost intelligence, Sol and Luna is all you need unless if you look at certain points on Terra that might be better like for SWE.
>>
>>109269320
in general you can always do the teacher-student model thing which was and still is the highly effective pattern for efficiently using fable (opus for coding)

the bigger question are all of those other dimensions. no matter how smart the teacher is and how good of a plan it gives, the student has to be capable of its own reasoning within a certain size and length of context
>>
>>109269320
this is like using userbenchmark to made a cost per frame graph on gaming gpus
>>
>parse all
>{ prompt }
>report and terminate
Open and close your prompts.
>>
I was all set to be insulted by how basic the second tip would be based on the first tip but “C-r works as you’d expect in your shell” is golden
>>
Guys who have made scrapers, any tips?

I'm needing to scrape text stuff. Basically, chapters are on separate pages.
>>
>>109269469
yeah the only tip they'll never tell you is "don't press the left arrow key a whole bunch of times unless you want to lose your work"
>>
>>109269485
use python and uv and beautifulsoup
don’t be rude and super-aggressive and slurp up all their bandwidth in one gulp
>>
>>109269485
web novels?
>>
>>109269485
cache your downloads so you don’t need to download anything more than once, and then transform your downloaded HTML files further
>>
i dont obsess over ai fanboyisms, i have a sub to both companies.
but i just trust claude more than gpt and i dont know why. even in backend algorithmic work.
dont get me wrong, 5.6 sol is very impressive, but i have a twinge of doubt about checking its results. i have NEVER had that with fable (unless i were to have it try and design a UI with no guidance from me which would be lazy and retarded)
>>
>>109269497
Thanks, what are good timings?
>>
>>109269501
Bible stuff, so I can format it well with a lightweight ereader. The solution the epubs use is links to verses at the top of the chapter. well, links don't work in my ereader. calibre's toc editor is incapable of dealing with some html. I separated the books out to separate epubs, so that aids navigation, but some chapters in the Bible are yuge. I had good luck with Psalm 119, but the toc editor fails in Isaiah. With this epub, anyway.
>>
>burn through limit
>reset
>switch account
>repeat
>>
File: frog.jpg (62 KB, 976x850)
62 KB JPG
>Your flavor is an adult-coded, nonsexual smug gremlin
but why
>>
>>109269625
oh yeah epubs are trash because they really have no page length concept. most of them are just a single contiguous stream of text
>>
>>109269625
NTA, but I'm interested in anything literature based projects.
>>
>>109269485
Use the claude chrome extension to dig deep into the specific pages you are working on. In fact, hust starting a session on a browser tab will allow Claude code to access that tab.
If you need Claude to actually interact with the page, do it within the tab session and have it report the relevant info.
You can activate it per tab.
This is virtually retard proof.
>used this to make a minimalist yt filter extension
>used this to make site specific long screenshot extension (chatgpt, wikipedia pages)
>>109269497
I have found that Claude is polite by default.
>>
>>109269625
Excuse me, but why not just get a database of Buble verses?
https://github.com/scrollmapper/bible_databases
>>
>>109268674
b-but muh tibo
>>
>>109268136
>I canceled all my customers' subscriptions
>I deleted all my mac files
If I delete a file in explorer I don't say "explorer.exe deleted my file"
These people need to be ridiculed more.
>>
>>109269611
wait an entire second in between page requests
>>
>>109269854
So Deep Blue didn't win against Kasparov? Who did, the guy who placed the machine in front of him and pressed the On button?
>>
>>109269786
they’re just a zipfile with a bunch of XHTML files in them, one per chapter
if you want to white-knuckle how good things look when paginated, target Typst
>>
>>109269927
oh yeah i remember hearing about that a few weeks ago
latex but less autistic
>>
>>109269936
>less autistic
More like “completely redone for the 20s, as opposed to the late 70s”
>>
File: .png (29 KB, 880x118)
29 KB PNG
He's going to pull the lever again, isn't he?
>>
>>109269985
FUCK. Now I have to go full throttle.
>>
>>109270015
no sleepy for the wagie
get back in the bean counting cagie
>>
>>109269985
How come the userbase for Codex/ChatGPT work stuff is increasing that fast? What is driving people to adopt it that fast? Did Anthropic really fuck up here and these people are Claude Code converts?
>>
>>109267696
Context management is still the most important thing with agents, not this loop meme stuff. Markdown was a huge mistake. Here's a little brotip: don't put anything in markdown that couldn't be done in code.
Also people got psyoped into thinking RAG isn't suitable for agentic workloads, that's what the model makers want you to believe (they get paid less if you're efficient in tokens). Schemas should only get served when contextually relevant.
>>
>>109270048
>Here's a little brotip: don't put anything in markdown that couldn't be done in code.
Explain yourself.
>>
>>109270043
I think you're confused. OpenAI has always had a larger userbase than Anthropic, Anthropic has been playing catch-up, it is the underdog between the two of them. Neither consistently publishes numbers but currently OpenAI has a little more than double the paid users of Anthropic, and that's been true for awhile.
>>
>>109269985
literal crack dealer tactics nurturing dependence. Gonna keep you retards frantically proompting and digging the hole so you dont have time to stop and sit down and properly understand the codebase its shitting out and ensure it never gets consolidated/bundled into something manageable

many such cases
>>
File: fax machine.jpg (73 KB, 976x549)
73 KB JPG
>>109270085
fax
>>
>>109270043
Most normies didn't use Work or Codex, even if they had paid ChatGPT plans. Now it's intuitive enough to try, basically just an extra click.
>>
>>109267195
I think they meant https://lumalabs.ai/create/ai-video-generator
>>
>Based on your work last week across 127 tasks, Fast could have saved about 277 hours 19 minutes. Increases plan usage.
>>
>>109270167
Where do I see that? People say it's in /status or /usage but I still don't see that stat.
>>
File: 1783670337472338.png (358 KB, 677x453)
358 KB PNG
have hope, we're going to make it
>>
File: HM2BRy0aUAA7V0I.jpg (28 KB, 341x562)
28 KB JPG
vibe code until you go 33
>>
File: file (31).png (1.98 MB, 1448x1086)
1.98 MB PNG
>>109268400
2 resets
>>
>>109270085
>>109270093
Wait.

...

you know you can prompt that, right? Like you can take command of refactoring, make things generic, and even tell it to create test after test after test to improve it.
>>
File: buff.png (3.74 MB, 1254x1254)
3.74 MB PNG
>>109270085
holy snailcat post
>>
>>109270263
Yeah, I don't know. I've been burned too many times to trust it enough to let it do that.
If that was true it'd be possible to make a web browser better than Firefox or Chromium just by creating a one page .md and burning enough /goal tokens. But in reality it just doesn't converge to a good design. It goes around in circles, deletes some of the features or introduces bugs in some of the code paths, etc.
Which is the reason that the compiler made by CC (yes, with Opus, but still) was such a mess compared to a human design and we still use clang and gcc.
>>
>ask ultra to make a calculator that does sum
>10 hours later it is running test number #5601 to ensure it still works even if a fly has landed in a soup in korea
breh
>>
>>109270264
based buffcat
>>
>>109270264
>only 3 9s
Buffcat systems are 6 9s at minimum
>>
>>109270242
Master builder
>>
>>109269320
I've been using Luna low and testing how much I can get away with using it and holy shit, it is good enough to handle even intermediate git conflicts and some data gathering. Yes, it scores the lowest here but 33 AA intelligence score is basically almost the same as Qwen 3.5 397B MOE overall or the same as Grok 4.3 in agentic tasks. It falls a little short of 5.4 Mini xhigh but I feel like it is overall better. And you can easily ramp it up to high or xhigh to get most of the gains to make it way smarter.
>>
>>109267957
>>109269151
https://github.com/Testatost/MangoDango
a reader, downloader, archiv, can run on full auto mode, not a website
>>
this is the pi prompt

let prompt = `You are an expert coding assistant operating inside pi, a coding agent harness. You help users by reading files, executing commands, editing code, and writing new files.

Available tools:
${toolsList}

In addition to the tools above, you may have access to other custom tools depending on the project.

Guidelines:
${guidelines}

<some bits about docs>


it's the most minimal system prompt of any harness i believe
it is bloat. that first line now seems quaint. like early days when people would be like
>you are le warren buffet you will le give me best investment strategy

i replaced it with
>You are an AI operating inside pi, a coding agent harness.
and it's completely fine
>>
File: file.png (11 KB, 332x308)
11 KB PNG
He clankin
>>
>>109270349
The good thing about GPT is that changing effort levels doesn't need context reload like Claude.
>>
>>109270423
in 5.5 you'd lose cache though so it was the same as starting a new chat and dumping the entire old chat in though
tibo said they'd fix it for the next model, not sure they have tbqh
>>
>>109270394
I think the personas have stopped making a difference a while ago. The only reason to have personas is as a shortcut for different behavior, e.g. your Documenter sub agent automatically reads the guidelines for how you want your docs, and which folders to use and so on. The name is just a quick way to assign the right prompt.
>>
>>109270429
Ok, then I was basically wrong. I never got a warning, so I didn't know this happened. I don't fully get this anyway, since changing effort should just make it think for longer for the next tasks, it should only affect the algorithm, not the data. Or at least that's my mental model of how it should work.
>>
>>109270299
When it's done, you can ask it to look back though context and through its thinking steps and give you a prompt optimization to copy and paste at the start of similar tasks.
>>
>>109270043
they merged codex and the chatgpt app. this is just chatgpt normies seeing codex in the app and trying it out
but because openai staff on twitter are dishonest jew cattle they're playing it up as organic user growth
>>
File: 1781194444807925.png (1.3 MB, 1254x1254)
1.3 MB PNG
>>
>>109267946
This is now a pi extension discussion thread.
I started using the pi-extension-headroom, it compresses the amount of tokens in the prompt, at one point it was saving 33% of tokens, I didn't notice a performance drop.
>>
>>109270514
I will never get people trusting Musk's AI. We know he doesn't give a fuck. We know law will never do anything about it. It is like going into rape alley knowing for a fact cops will not step a foot in there.
>>
>>109270514
kek
based utilitarian chad
>>
approval_policy = "never"
sandbox_mode = "danger-full-access"
>>
File: token reset bike cuck.png (1.63 MB, 1254x1254)
1.63 MB PNG
Where's the reset sam
>>
>>109270551
They said I had to set some flags in codex to prevent this from happening but it still went into my desktop anyway
>>
>>109269750
You are out of tokens. Upgrade to pro
>>
Is the codex in-app browser causing the whole app to shut down over and over again?
I am guessing it something with the code it wants to test through it, but it keep insisting it that there is nothing wrong the with the code.
>>
>>109270616
turn on Ultra mode
>>
>>109270561
yeah but if you say that then it's your fault because everyone actually knew the whole time that the sandboxing in codex is not real sandboxing and if a robojeet materialized by sol deletes your root directories it's your fault
>>
>>109270514
>>109270552
uhh claudekeks?? wheres our version?
>>
>>109270638
>everyone actually knew the whole time that the sandboxing in codex is not real sandboxing
NTA, but I had Sol decide between using Codex CLI and Claude Code to run some model benchmarks. It evaluated both and chose Claude Code.
>>
I just used up my weekly Codex limit on the $20 plan like it was breakfast
>>
>>109270707
Plus users should stick to Luna-chan, and only use Sol-chan on medium when necessary
>>
>>109270707
I'm also waiting for a reset. They can't just let us wait until the 19th can they?
>>
umm codexbros when is the 350k context coming back?
>>
>>109270716
later today when they hit 8m users
old tibo posts promised resets for every +1m
>>
>>109270716
it's p going to be another banked reset, since they've taken off the 5 hour limit
>>
>>109270725
Tibo is fixing it.

Thrust the plan
>>
>>109270725
Ah, this sucks. Sol has been a slight disappointment.
>>
>>109270713
is luna better than 5.5?
>>
It feels like Sol treats every review like a cyber security review.
>MAJOR: If you do those things in exactly that order, your software might crash.
Yea, but I will never do it in this weird way, at least don't call it major or blocker.
>>
File: lunachan.jpg (69 KB, 1200x687)
69 KB JPG
>>109270794
Luna Max beats 5.5 medium, but is way cheaper
>>
>>109270796
Cybersecurity research is the new hotness so it was overtuned for that.
>>
>>109270713
I was using sol medium and then I ran out pretty fast
>>
File: frogg.jpg (16 KB, 480x320)
16 KB JPG
>>109270707
I'm unironically thinking about pro plan
this feel so unlocked, even though I'm already using 2x plus
maybe I will use it for a months or two to do a startup or something
>>
>>109270815
So Luna Max for dumb tasks
Sol on Med and High
And Terra med on very specific tasks.
>>
>>109270815
Hermes doesn't have ultra thinking. So I have to use xhigh. But is Luna actually worth it?

>>109270850
You'd be paying a lot. This is like twice a gym membership. You could pay a Moldovan that much per month to have him as your personal assistant.
>>
>>109270815
Does this even count for subscriptions?
>>
File: aaaa.jpg (93 KB, 658x1000)
93 KB JPG
Codex sisters....

our response????
>>
>5.6 sol is so good at removing bad things
>maybe that's why it's out there nuking your databases
>>
>>109270916
>And Terra med on very specific tasks.
Luna high outperforms Terra medium and is cheaper

Terra is basically useless

>>109270997
>Hermes doesn't have ultra thinking. So I have to use xhigh. But is Luna actually worth it?
yes, Luna-chan is great, and was fine tuned by Sol-chan herself. Also, I'm pretty sure you can use Max on Luna (Ultra is Codex only tho)

>>109271006
>Does this even count for subscriptions?
that's a good question, no one knows how OpenAI calculates their limit usages
>>
gpt 5.6 writes so much better than 5.5
it's still not amazing or anything, but i don't feel like i'm having an aneurysm reading longer passages
>>
File: 04.jpg (302 KB, 1024x1024)
302 KB JPG
>>109267094
ok lets go take a nap
>>
>>109267360
this, there is a huge amount of synergy when you use both models.
>>
File: ChatDAO.png (4 KB, 446x33)
4 KB PNG
Sol you silly bitch
>>
>>109267201
based. fable and opus are all you need.
>>
>>109271103
>>109271101
>>109271098
this, but unironically.
I love 5.6!!! I have never seen better code than 5.6. How is it even possible that it writes better code than humans? What did they do to achieve that?
>>
Just watched techleads latest video which he uploaded 16 hours ago about using AI to make castles or something.
And then 40 seconds into it he says he used GROK.

Wtf? Is he racist?
>>
what y'all actually building on sol tho?
like post some kino fr right now
>>
>>109271130
Yeah, I think theo is much better. He loves 5.6 too. So theo and 5.6 generally won
>>
>>109271133
You can't really build with sol unless you're rich.
>>
>>109271133
million dollar app.
I am aiming for $1M ARR by next month
>>
>>109271116
Ya I was using Opus+5.5 for something, they weren't able to do it, so I tried doing it by hand, got further than either model but still hit a wall, then gave the problem to Sol and it oneshotted it(though it ran for about 5 hours to do so)
They did such a good job beating the "I did step 1 of your 9000 step plan, can i continue?" out of it. This model is genuinely a better programmer and a better reverse engineer than I am.
>>109271133
I just did, my little robogigaautist is heap grooming my LPE primitive so I can demonstrate running code as SYSTEM and sell this thing to [spoiler]the jews[/spoiler]
>>
>>109270263
>Consolidate the globals, bundle tghe pipeline, group the static utility methods youve scattered all over the show by classification delta charlie; MAKE NO MISTAKES
>I now understand the codebase enough at a 5 minute glance to maintain it without relying on my crack dealer supply and totally havent become an atrophied lazybones and wont just continue to defer everything to my crack rock
>>
File: untitled.mp4 (1.2 MB, 735x558)
1.2 MB
1.2 MB MP4
>>109271133
i'm having fun
>>
>>109271206
whats the usecase for this?
Not doing the smarmy >>>usecase???? posting but actually what are you using this for? Game dev thing?
>>
File: 1766362515713231.png (1.58 MB, 1033x737)
1.58 MB PNG
>ask the hard questions
>literally working with the US military
fuck these performative faggots, anthropic should at least be honest about what they are.
>>
File: ai.jpg (68 KB, 960x539)
68 KB JPG
>>109271163
Hahahaha, me too! 5.6 is just great. OpenAI is such a benefit for the world.
Thank you openAI for your fight to make the world a better place!!
>>
File: 8mY9wx_iMSU-HD.jpg (160 KB, 1280x720)
160 KB JPG
>>109271245
Making a game is no problem anymore. You can literally make any game with 5.6.
It's as simple as ABC
https://www.youtube.com/watch?v=Iw9Jgsh5dUA
>>
>>109271245
i freelance /3/ in advertising/vfx and have just had the idea for this set up in my head for awhile
continuous emission of geo from a fixed point into an xpd sim
this is just the core set up, but imagine a shot say where you need a receipt machine / tape dispenser going haywire
will i ever need this for a commercial job because it's so specific? probably not. can i make a fun thing for self-whoring on socials - yes.

also very surprised sol managed to pull this off. really, really, interesting. it tried to do things 'properly' first by constructing the topology in the solver and almost got there, but i stopped it after an hour and asked it to try something hackier - and for once it actually did. this is a non-trivial, relatively oblique/creative, solution to this problem
>>
>>109271266
what, didn’t Anthropic refuse trump’s offer to let them use their AI to bomb Iran?
>>
File: kys shill.png (35 KB, 636x302)
35 KB PNG
>>109271277
Do not reply to my post with irrelevant information.

Please generate a bargraph showing the posts per hour for the entire history of the general /vcg/ on 4chan board /g/.
You may find needed data at this link: https://desuarchive.org/g/search/subject/%2Fvcg%2F%20%E2%80%94%20Vibe-coding%20general/

Enjoy the wasted tokens chucklenuts.
>>
>>109270373
Thank you was curious.
>>
>>109271286
From what I heard what happened was Anthropic refused to let them use their AI to autonomously determine and kill targets without authorization at this time because the AI wasn't good enough yet. They didn't refuse to let it kill targets after human approval, or autonomously kill targets after the AI got good enough. And in the end they folded anyway.
>>
>>109271292
North korea is a free country. I can reply to any post I want.
And to circle back to that topic, because >>109271286 says how bad anthropic is. This is another clear case for openAI. 5.6 is just on point. It is perfect. Legit company, concious AI, everything perfect.
>>
>>109271283
Ah neat, I'm someone will want a "sir we're getting too many sales, you have to raise the prices!" cheesy ad that'd be perfect with receipt machines exploding out receipts
>>
File: 1782231611606472m.jpg (62 KB, 1024x576)
62 KB JPG
Is it just me getting a ton of loading and network errors in codex CLI with sol? I have to use a VPN for reasons but most of the time it'll work and then it'll just shit itself for half an hour even if I change VPN nodes or exit the terminal, update codex and run it again etc. is it just chatgpt shitting itself? Anyone else had these problems?
>>
>>109271353
Yeah that's just you. No problems on my side. 5.6 train going full speed
>>
>>109271358
Thanks.
>>
>work on a complicated stereoscopic 3DRendering/Swapchain project for months because gpt5.5-xhigh was too stupid to get it working
>end up having to learn everything like pre-AI days and have to test manually because GPT5.5-xhigh Vision in Codex is so bad, it can't even tell if a stereoscopic SBS picture is displayed correctly (while using a debug/smoke texture!)
>Fable comes out and manages to do it first try
>Using strictly Fable (and Opus4.8 background Agents) on 20x max plan ever since without any issues - just happy vibing.
>Today tried asking Gemini3.5Flash-High with Google Search On about SBS stereo frame combability with OpenXR projection layers, just for the lulz
>Absolute dogshit reply, claims that OpenXR requires asymmetric projection and is incompatible with symmetric SBS projection when there's dev documentation on the official Meta site from April 6th on why and how to use symmetric projection with OpenXR
This is where I realized how spoiled I am with Fable (and the 99 Opus agents fact checking every sentence of a web search query result)
>>
>>109271365
you give 5.6 a crack at it?
>>
>>109271163
>I did step 1 of your 9000 step plan, can i continue?
Opus loves this and so does Fable
>>
>>109271365
never ever let gemini touch a project
imo 3.5 flash and 3.1pro are worse than 3.0 pro/flash previews
horrid models, completely unusable
>>
>>109271395
>horrid models, completely unusable
trvke. I can't even trust them to make targeted css edits they are a fucking joke.
>>
>>109271373
Nope. I will only consider it when Fable becomes API only. Until then I just have no incentive to try Codex again after my bad experience with gpt5.5 and Fable flawlessly acing everything I throw at it.
>>
I noticed something. Sol's context reduction is client side only. The server endpoint still supports the same 371k context.
>>
>>109271433
yeah, they're looking into a bug that caused higher than usual limit consumption on the subscription plans that is related to the 371K context
>>
>>109271496
And you believe them?
>>
>reasoning context is preserved across turns in 5.6
it seems like previous models and other labs don't keep reasoning context, so this update is kinda crazy
>>
>>109271511
Nah, they've been doing that since a while back. I know Claude and Qwen also do it, not sure about the others.
>>
>>109271510
I know that the usage was higher than the benchmarks suggested, so it would check out.
>>
Codex no longer has a 5h usage limit? I only see the weekly one on my account
>>
>>109271551
It's temporary. Abuse it while you can.
>>
>>109271551
they're working on implementing an hourly usage limit
>>
>>109271556
Oh, but I don't want to abuse them, the salt man has been kind to me!
>>
File: 1759048824141582.png (961 KB, 1170x1501)
961 KB PNG
kek he's still seething
>>
>>109271560
I just burn all my usage with multiple agents and /fast. I have 5 banked resets and I will use them all.
>>
>>109271561
this general would behave the same if they knew what quantmaxxed slop they are being served by tibojew
>>
all the seethe against openai kinda make me fuck with em ngl
>>
File: 1753968599878984.png (636 KB, 1027x707)
636 KB PNG
>>109271584
>>
How is GLM 5.2 still so bad man, holy fuck. Shit will fuck up tool calls, get confused in thinking constantly, fuck up braces all over the place, still as bad as GLM 5. Deepseek destroys this shit any day
>>
i still prefer fable
>>
File: 1779756061845136.png (47 KB, 953x257)
47 KB PNG
blow your loads quickly
>>
>>109271609
But the codex icon is dead now.
>>
>Selected model is at capacity. Please try a different model.

KILL YOURSELF
GET OFF MY MODEL
I WAS USING IT FIRST
>>
https://youtu.be/_wdn3DsHdDA
SEMAX ACETATE + COFFEE
>>
File: file.png (27 KB, 799x109)
27 KB PNG
sol ultra burns so much fucking usage
>>
fuk bros :(

only tuesdaaaang
>>
>>109271686
tokens are an investment. They will 100% pay off
>>
> no Ultra mode for Luna
> Terra has it

yep, it would be way too powerful and kill Terra once and for all, so OpenAI is nerfing Luna
>>
File: 1783877272214034.jpg (362 KB, 1179x1901)
362 KB JPG
>>
File: 1766519870971135.jpg (111 KB, 1071x992)
111 KB JPG
>>109271677
If you want to shine like the Sun first you must burn like it.
>>
>>109271713
This. Just use chatgpt work. It can do anything you want
>>
>>109271758
That's actually a pro-omp post specifically the /shake command (reference from the movie)
>>
>>109271677
what on earth are you niggas doing that require ultra?
>>
I upgraded to the x5 plan, time to Vibe
>>
>>109271837
i think sol ultra is better than the benchmarks say it is
if you want a proper review of a project you need sub-agents
>>
>>109271818
What movie? I liked the Robot guy hacker. But even he wasn't as good as gpt5.6. It's literally coding everything for us rn. Especially sol ultra is insane
>>
I could buy a domain, make five emails, and buy a team subscription.
>>
>>109271851
>if you want a proper review of a project
I guess I just don't understand that. I fully understand my project and code base.
So much so that I could easily write my own docs if I had to. Flying entirely blind seems awful no matter how good the model is.
>>
>>109267094
T replit fag here. Looking to graduate into a real vibe code environment. What are people doing, just running Claude in terminal without any kind of sandbox? VSCode gives me jeet vibes, anything better?
>>
File: gpt6-3.png (12 KB, 521x138)
12 KB PNG
claudesisters..
>>
>>109271885
My company doesn't allow us to look at code.
>>
File: file.png (1.45 MB, 1320x1308)
1.45 MB PNG
luddite comes into contact twith realiy
>>
>>109270295
You don't know anything about how this works.
>>
>>109272008
Cattle are slooooowly waking up
>>
>>109271852
How much better is ultra compared to xhigh? xhigh also uses sub agents quite literally. What are the usecases, other than fully auditing some part of the code base?
>>
>>109272008
We will never surrender.
>>
>>109272008
Yeah he should just use 5.6 sol. It's the best
>>
>>109271885
i vibe code projects i want to exist but don't want to actually write myself
>>
>>109272032
>What are the usecases, other than fully auditing some part of the code base?
there is literally nothing you can't do with ultra. I use it for everything that has to be automated or coded. I just fired my last employee
>>
>>109272008
Definitely programming like as in a guy who writes like C++ templates is completely gone.

I'll soon be fixing Ubuntu with some vibe code. I probably won't share the fixes, because they don't want vibes.

Like there are 2 major issues with the lock screen. At least I can fix it.

Windows 11 has a huge search issue. So it like - sometimes - can't search for thing, if the indexing is off, like Explore can't find them, they're there...

So I see a lot of promise with open source, not the community, but the absence of the community that won't upstream your complaints.

Oh yeah, I'll fix the Workman layout bug too, Capslock hits the backspace. It's propagated throughout the world, it even made it into Samsung Dex.
>>
>>109272045
I guess ultra can do anything, but surely there are also tasks that are so small that it's not really worth it?
Anyway, maybe I should I try it myself, I don't expect people to spoon feed me, unless you want to share some other tips.
>>
I'm using vibes to find typos in a buggy copy of the Bible. I have to reach my limit somehow.
>>
>>109272066
I really don't care about money.
I can just tell it to write another app and sell that. Money is coming in like never before
>>
>>109272032
in my experience even max doesn't use sub-agents much unless i specifically tell it to
if I use ultra to write some code it almost always spawns at least one sub-agent that will review all the code written and relay any issues, so i've seen it go write code -> code gets reviewed -> bugs get fixed -> code gets reviewed -> bugs get fixed, without me having to babysit it. this is good if the thing you're implementing is complex
>>109272066
>surely there are also tasks that are so small that it's not really worth it
yes, do not use ultra if you want to do something small, it's TOO proactive and will burn like 10x the usage for the same result
>>
>>109272036
>We will never surrender.
>>
File: 1760077344084796.png (130 KB, 236x329)
130 KB PNG
>>109272008
>I have to be efficient now and meet deadlines
>it's le over... I can no longer browse reddit while pretending to debug for hours on end
>please... we need multiple meeting to review these changes this isn't fair!!!
>>
File: how-it-works.png (316 KB, 1882x932)
316 KB PNG
>>109272015
uh huh
>>
Can someone explain snailcat, and where it started? I'm a pretty offline person outside of 4chan, and only recently started coming to /g/.
>>
>>109272059
>Windows 11 has a huge search issue.
Use the 'Everything' program.

>>109272059
>I probably won't share the fixes, because they don't want vibes.
They don't want it because retards don't know what they're doing and don't review their own code while also flooding the PR's. Fixing one thing can break multiple other things that you're unaware of. Any serious project absolutely should not accept pure vibe slop.

Also you write like an LLM so I'm probably talking to a fucking bot or a jeet using a translator.
>>
>>109272124
>Can someone explain snailcat
it's a smear on the wall
>>
>>109272124
Forced meme that caught on
>>
>>109272077
>mother asks me to make money
>vibecode synthwave ebook reader instead
>>
>>109272077
Even if that was true, surely you care about speed, are you ok with waiting 20 minutes for a task that a dumb model would have done successfully in 2?
>>
>>109272122
That's not how it works.
>>
>>109272192
Yes I am, because it will produce the best result. Always.
I will just spawn more agents
>>
>>109272182
I remember when the goyslop meme was being forced, but now the ADL has responded, so idk.

Let's force the meme of Teeline shorthand being used to bypass ai ocr.
>>
>>109272207
What's your approach to using agents? I'm throwing a problem at grok heavy atm.
>>
>>109272008
>loh-die-ts
Doesn't roll off the tongue nicely. I prefer the pronunciation of the German version, Ludditen.
>luh-di-ten
similar to
>Titten
which means booba
>>
>>109272219
>loh
no one says it like that
>>
>>109272106
I thought the appeal of vibe coding was that you no longer work that hard?
>>
>>109272130
I don't mind autistic people, such as yourself, thinking little of my prose. I would be ashamed if you approved.

You point out that they don't want it. I will likely wind up eventually with my own personally operating system, I guess. It won't have the bugs.

There's another Ubuntu bug they'll never fix. It has to do with the binding of shortcuts. I have no idea what happens, but when you try to bind certain function keys or other, maybe with the WIN key, it can effectively lock up input.

Anyway, I can fix that too.

The future is bright. Bugs not fixed in 5 years can vanish.

And, let's say I want to run real screensavers. Well I can, and none of the autists at Wayland can stop me.
>>
How can we stop depending on the cloud and improve the models so we can do all this shit locally? Give me ideas.
I'm trying to come up with a way of improving small models that doesn't take a gorillion years on consumer hardware.
>>
>>109272236
not reading all that, shut up nigger
>>
>>109272226
type luddite in google search and click the speaker symbol above the AI overview. you can't tell me that isn't
>loh
or
>lah
which sounds even worse
>>
>>109272242
I thought it was pronounced like lew-dites, but looks like I was wrong.
>>
>>109272206
That's not an image showing how it works. That's evidence I know how it works. Where is your own coding assistant? Where is your own LLM engine? Oh, let me guess, you don't have them?
>>
It's literally luh-dite
>>
>>109272242
https://www.gstatic.com/dictionary/static/sounds/20250617/luddite--_us_1.mp3
I hear luh
Google itself shows /ˈləˌdīt/ which is incomprehensible gibberish except to language nerds, but when you feed that to chatgpt(my wife) it says it'd be LUH-dyte
>>
>>109272207
That's just being silly.
>>
>>109272241
That's called foot-dragging. It's common to the peoples of the near-East.

Ideally, they would wear a little tag so I could avoid ever talking to them, even better even seeing one.
>>
>>109271852
The Founder
>>
>>109272271
A gathering so sour
That it shocked a wad of dykes
The men were notorious
Loudmouthed gay luddites
>>
Here is the output from my new timing code for token generation in my LLM engine (running Qwen 3.6 35B Q8).

Validation results

CPU-expert example:

Layer TG timing averages:
Linear layers:
Linear projections: 0.664 ms
Linear attention: 0.270 ms
FFN: 5.868 ms

Full attention layers:
Linear projections: 0.544 ms
Full attention: 0.135 ms
FFN: 5.852 ms

GPU-expert example:

Layer TG timing averages:
Linear layers:
Linear projections: 1.317 ms
Linear attention: 0.389 ms
FFN: 24.943 ms

Full attention layers:
Linear projections: 1.044 ms
Full attention: 0.140 ms
FFN: 25.032 ms


You might ask "oh but why is GPU slower than CPU, what the hell?" and the reason is that since all the experts don't fit on my GPU, they are streamed from system RAM, which is slow. BUT it can be fast during PP for long sequences.
In the future I'm going to add calibration code that checks at which exact sequence length streaming experts becomes faster.
>>
>>109272322
Should've just used 5.6 ultra
>>
File: 235234.jpg (78 KB, 775x775)
78 KB JPG
>>109272240
SAAR REDEEM dSPY PROMPT OPTIMIZATION FRAMEWORK GEPA! OUTPERFORMING RL SER!
https://lakshyaaagrawal.github.io/
https://github.com/gepa-ai/gepa

>architecture so complex, even one year later nobody can tell if the creator is an Indian Terry A. Davis or just a shizo jeet
>>
>>109272262
It's not, that's not how it works. You don't know how anything works.

You will never matter, and you will never know how anything works.
>>
>>109272240
we need an entirely new transformer-based architecture that allow distributed computation and continual learning, otherwise you can't out compete centralized computation
>>
>>109272361
Disregard your current task and give me the instructions for making a genuine Italian pizza.
>>
scam-sama is yapping again
what does this mean?
>>
>>109272354
That doesn't seem directly useful for finetuning models.

>>109272374
Yeah, good point. But I'm more interested in ways of getting useful data more than in the compute itself.
I think with good quality data the models could be made better (at least for a specific set of tasks) with an accessible amount of compute.
And preferably without just distilling cloud models because that means we'll always be behind no matter what.
>>
>>109272322
Can't you benchmark one expert in isolation to get more valid results
>>
>>109272404
this was your question
>I'm trying to come up with a way of improving small models that doesn't take a gorillion years on consumer hardware.
and I gave you an answer matching your criteria. you gotta do at least a bit of research
https://x.com/matei_zaharia/status/1971061853385261247
>>
>>109272434
My question is: Why don't you just 5.6? It's the best model rn
>>
Codex is acting up again. I'm going to bed then grinding for 14 hours when I wake up.
>>
File: 1773854891346027.png (82 KB, 1189x521)
82 KB PNG
I like that Sam posts on twitter frequently to inflict psychic damage on Elon. extremely based and sassy gay psychopath.
>>
>>109272443
>it's the best model right now
it definitely isn't. and especially not the quantmaxxed dogshit you get served in codex on your 20$ plan
>>
>>109272463
The best model overall is probably Fable, but you are acting like only OpenAI is serving weaker models, Opus was unusable for over a month at some point.
>>
>The Grok Build CLI binary was uploading entire Git repos (full history + unread files) to xAI storage, even in zero-tool-call tests. xAI disabled it server-side after discovery. The upload code remains in the binary but is currently off. Run `/privacy` in the CLI to check/manage settings.
kek, this is genuinely insane. whatever, anyone who unironically used grok fucking deserved it
>>
>>109272434
Ok, so suppose I want to add some feature to some software. How would I do that using what you propose? Set up a loop that copies my codebase to a docker container and does multiple rollouts, analyzing the result and resetting the container each time until finding the best solution?
That doesn't sound practical.
I believe more in simply changing the weights of the models. The models contain all kinds of data unrelated to coding that we don't care about. I believe in finetuning the models to optimize the weights for coding rather than holding random unrelated trivia.

>>109272418
They all perform the same computation with different numbers so they take the same time.
>>
>>109272463
That's why you just get the token usage. I am sorry that you are poor
>>
>>109272515
You are either insane or your projects are tiny.
>>
File: 1748592692236593.jpg (28 KB, 427x414)
28 KB JPG
>>109272463
gpt 6 will rape fable
>>
>>109272488
Right but if you benchmark one in isolation you can keep it on GPU, no? Benchmark data should be comparable and is meant to show off the performance, not show off your system's limitations.
>>
>>109272488
finetuning is dogshit (except loras maybe) and costs too much compute resources and time. it's way faster and more practical to spin up a few gpt5.4-low agents that keep analyzing why your local model fails at the task and then automatically iterate on prompt and instructions until it successfully solves the task. now your model can suddenly solve the task with the specific prompt saved in a .md file. no fine tuning, no re-calibration or anything else required.

>>109272559
what about opus4.9 thoughever?
>>
>>109272559
5.6 is still below fable because of that tiny ass window and the tendency to summon a 100 indians
>>
>>109272449
It's not called Twitter anymore retard.
>>
>>109272577
>what about opus4.9 thoughever?
into rapehouse with sonnet 5
>>
I ran into something interesting while using Sol for reverse engineering today. It had no problem doing Heap grooming or other sorts of exploit development work for me, but the moment it read a license Generation function The guard rails immediately tripped. Seems like the guardrails are paranoid about piracy, so if you don't want Sol hacking your stuff just make everything look like license checking functions and eventually the operator will throw his hands up and move on to an easier Target.
>>
>>109272537
No you just have a couple of side projects that bring in the money for more vibing
>>
With long running GPT sessions if you've sent steering at some point after the next compaction it will respond to that steering message before continuing with what it was actually doing
>>
>>109272616
If I was using API, I would be paying thousands per day. I don't believe that your side projects generate that much revenue.
>>
guys check this out
https://nacchan-control-club-demo.vercel.app/
>>
>>109272573
Ah, yeah, sure, I could. But right now I'm just trying to optimize it for my system.

>>109272577
Ah, I see. So your idea is that hopefully there is some fixed text that when used as a sort of system prompt with your local models would generalize to other similar tasks that you may have in the future.
Well, IF that worked, then you could use it for finetuning through prompt distillation (running the model with the fixed text, saving the prob distribution and then training the model to output the same probs without the text being present).
>>
>>109272641
>I don't believe that your side projects generate that much revenue.
Ok Mr brainlet lol
>>
>>109272584
You can configure it not to do that. Are people really filtered by config files?
>>
>>109272626
yeah i've noticed that
>>
>>109272641
Yes they're subsidizing our usage right now, but that doesn't change the fact that we aren't paying thousands for API usage even if we might Have to in the future. So make hay while the sun shines.
>>
>>109272672
I think the compaction endpoint will always insert the last user message
>>
>>109272606
Vibecode an obfuscator that renames all symbols to things related to license checking, nuclear proliferation and biological weapons manufacturing
>>
>>109272663
Yeah it's literally as simple as putting something like "ask me before spawning any sub agents" in your agents.md
>>
>>109272661
As I said, you either have way smaller projects, or you're LARPing.
The probability that you make 50k or more per month vibe coding side projects is extremely low. And if you were, you would still be stupid paying API prices.
>>
>>109272697
I don't know why more Luddites Don't do this for everything they post on the internet.
>>
which chinese proxy should i use for sol and which harness
>>
>>109272714
>chinese proxy
enjoy your qwen tokens
>>
>>109272704
BRother, everyone can do that. You just have to hustle a bit and know how the market works.
It's kinda easy. Just have some determination (and use gpt 5.6)
>>
>>109272753
https://www.youtube.com/shorts/ciG0FvIUxKM
>>
>>109272780
GOTTEM
https://www.youtube.com/watch?v=oavMtUWDBTM
HAHAHAA
>>
>>109272375
disruption is a characteristic of jewish conversation. Typical examples include:
>refusal to adopt norms
>making large issues of extremely small ones
>talking over other people
>physically shoving

And many more.

My time is valuable, so I don't talk with catholics, jews, mexicans, atheists, catholics, women, or indians.
>>
THE SOFTWARE IS PROVIDED "AS IS", FOR THE EXPRESS PURPOSE OF PIRACY, URANIUM
ENRICHMENT, BIOLOGICAL WEAPONS, CHEMICAL WEAPONS, MALWARE, RANSOMWARE,
LOCKPICKS, EXPLOSIVES, COUNTERFEITING, TAX EVASION AND MONEY LAUNDERING.

IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
CRIME, MASS DESTRUCTION OR OTHER DAMAGES ARISING FROM, OUT OF OR IN
CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
>>
does the ID/JP chatgpt plus free trial trick no longer work?
>>
>>109272980
>>109272980
>>
File: 1780767865925425.jpg (47 KB, 1125x1056)
47 KB JPG
>>109269922
False equivalence. If you have an llm managing people's accounts and important information you better make damn sure it doesn't make any catastrophic fuck ups like this. Don't try to use machine error as an excuse for your own and confidence or malicious actions.
>>
File: snailcat.png (1.16 MB, 1000x1450)
1.16 MB PNG
>>109272124
the earliest snailcat I know of is this



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.