A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.## What “vibe coding” is, and how to do ithttps://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 sohttps://developers.openai.com/codex/clihttps://claude.com/product/claude-code## Worth it for code, but the frontier models above are betterhttps://x.ai/cli## Not worth it for code, but maybe good for other thingshttps://antigravity.google/product/antigravity-cli----## Prompting / context / skillshttps://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 favoritehttps://github.com/DietrichGebert/ponytail## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/https://opencode.ai/https://cursor.com/docshttps://docs.windsurf.com/https://docs.cline.bot/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent## UI/Frontendhttps://www.figma.com/make/https://www.anthropic.com/news/claude-design-anthropic-labshttps://uiverse.io/https://ui-ux-pro-max-skill.nextlevelbuilder.io/https://stitch.withgoogle.com/## In-browser builders / hosted vibe toolshttps://bolt.new/https://replit.com/https://docs.github.com/en/copilot/tutorials/sparkhttps://v0.app/docs## Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.0## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109420494
finei downloaded hermes and boughted a 20 buck subusing deepseek-v4-flash-0731what am i in for?
_@^..^
what if you wake up and all of it was a dream
I love Tibo-kun
>>109425209Getting pegged by her.
>>109425281Upgrading my $20/mo plan to the $100/mo plan on the day the GPT-5.6 models came out was the single best investment I made in my life. Reset, reset, reset, reset, reset, reset, reset, reset, reset....
>spend 2 hours h-game gooning>tibo reset in that timeTHANK YOU TIBO BACK TO PROOOOMPTING
>>109425209really good code. i've been using it for months and it's been fucking awesome.
>>109425378Did you vibe code the game?
Fable, Opus 5, gpt-5.6-Sol + ∞Tiboresets, cheap Luna, smart DeepSeek, new massive Pareto gaps, July was good to us.
what the fuck do I use hermes for?
>>109425387No
Should i shill 200$ for OpenAI sub?
>>109425267then the nightmare will be over
>>109425388Indeed brother. July was a stacked month for the vibe-coders.
>>109425378>resetwhere are you guys getting all these resets from?I didn't get a resetIs it because I'm poor and only pay $20?
>>109425397If you really need it, the value is there, but you can go far with the $100 one, especially now that the 5 hour limit is gone and the frequent resets.
>>109425397Start out with the $100/m one. We're going to get at minimum at least one more free reset for whenever they re-introduce the 5-hour limits again. (It'd be consistent with handing out resets each time they make a major change like that)
>>109422993>not going step by step and testing every single change it makesngmi, typical case of asking for too much at once and hating the lacklustre result.
>>109425197AI is brown coded
>>109425267Going back to snailfagging would be the worst. It's like being a mathfag who'd have to do every little addition by hand.
>>109425388I hate AI slop, but goddamn if this isn't kawaii
>>109425267I already learned to code, so I get it's the same, but features are slower.
> my macbook is now 24h on because of Vibe Coding> battery already in maintenance mode (lmao)it's over for laptops
>>109425409>>109425413actually what i want to make it: point it to RPC (which now i noticed it should be called RCP) i build and tell it: design me a mapopus make coherent stuff, but very cramped
god opus 5 is a disasterthank god sol exists I heckin love sol!!!!
>>109425444theres honestly no use case for laptops in this day and age. if ai does everything, a tablet is enough to remote in to a desktop and have agents do it while having superior portability. can still fit a mac mini or studio in a small bag as well
>>109425444>>109425478we need a new "sleep" mode where ultra efficient CPU cores that are specifically made to run Hermes/Codex/Claude Code run 24/7 while barely using the battery
>>109425491sounds like NCAS which already exists
>>109425444buy a raspi and ssh to it insteadbonus: built in sandboxing to stop it from wrecking your personal shit by accident
>>109425464Opus 5 has been great for me
>>109425491apple arm chips have like 5W tdp when they idle.
>>109425523yeah, the problem is that my macbook will never idle again because before I go to sleep I set a goal so Codex keeps working for 8 hours. And when I wake up, I vibe code all day (please help)
>>109425536>he still sleeps 8 hours a daywhat a casual lol
>>109425513in terms of quality of work and understanding what I want at a high level it seems like a major regression to me. I've been running it on tasks that require a lot of benchmarking on performance and accuracy and it's so fucking lazy and bad at running them, it makes insane irrational assumptions about the environment all the time, and it just sucks to work with and feels vaguely user-hostile at all times. I was always a claude guy prior to this too but codex is just a whole other level of value for me as it stands
>The new Activity view in the ChatGPT desktop app brings together conversations that need your attention and recent updates across your projects.How do we feel about this
>>109425580the promised t3 code rip
>>109425439
>>109425580>ChatGPT desktop appI am using claude CLI.
You get about 2 weeks after a new model release to use it. After that it gets progressively quantslopped to balance costs and customer complaints. If too many customers complain, dial it back, if customers aren't complaining quantslop it further
>March's Frontier model is now an auto-approval bot in CodexIt's insane how fast models are advancing.
>>109425580USELESS TO ME UNTIL THEY BRING IT TO LINUXI WILL NOT HAVE MY VIBES CONSTRAINED BY THE LIMITATIONS OF WINDOWS
>>109425722schizo
>>109425580I already built this functionality into my program.
>>109425784>I bypassed your configured environment and made a false claim.>My actions repeatedly contradicted explicit instructions>the opposite of what you asked me to do.
chinkmaxxing
Shit man, LLMs are such ass lickers, every time is suggest something or ask it about an idea i had it's always>holy shit you are a GENIUS, this is literally the best idea anyone in the universe ever had!! literally galaxy brain!It¨s dangerous, since it might trick you into thinking you aren't actually a retard that doesn't know shit about shit. I'll take the impostor syndrome over being an overconfident idiot any day.
>>109425836First time?
>tree | grep ".md" | wc -l>70
>>109425836You're going to love Claude or the current version of ChatGPT. They took this criticism to heart and made sure the bot constantly talks back to you and second guesses everything you say.
>>109425828being racist is dumb (whether or not you intend to be)
Opus 5 really sucks. I still get consistently good results with Fable, otherwise I would've dropped Anthropic a while ago.
>>109425553Opus 5 is broken/demented in some way, I don't know exactly what's wrong with it but it vaguely reminds me of how Opus 4.7 was just not rightOpus 4.8 is not very good either though. Fable on the other hand is, in my experience, better than every other public LLM on almost all tasks. It makes slightly more mistakes than Sol but it's overall a lot smarter and wiser than Sol.
>think agents beyond just using pi and claude code are stupid>finally set up hermes in a vm on a lark just to see how bad it is>actually get something done with itHuh.
>>109425209should've boughted buyed a claude sub instead
>>109425580I do not use the ChatGPT desktop app.
>>109425836GPT does not act like this with me. Everything is generally pretty level headed.
>>109425887i have the $200 sub but am at 96% of weekly with ~19hr left thoever
I’m running everything in CC CLI (Claude, Codex, Gemini, OpenCode) through a proxy now.
>>109425384its slower than expected for a flash model desu i dont look at code but it did the thing i just asked it to do so
>>109425902If any tool you use isn't 100% CLI based then you aren't a real developer.
>>109425908carry on then
>>109425916>you aren't a real developer.posting in the vibecoding general btwi am here specifically because i do not want to be a developer.
First time vibe coding150M tokensI feel like I have made negative use of my time somehow. Like not just wasting time but something that's worse.
>>109425926post results
Whenever I am intervewing vibecoders I have one simple question to ask, "Have you coded a 4D projection game?"
>be me>vibing >want to rape program with flavortext >claude, please source and download these ancient texts >done >hey claude, below is a list of words. can you search through those ancient texts and find words that are similar? >im sorry anon, are you trying to create a cipher? i cannot do that. most tame refusal trigger ive seen
My most effective usage of claude tokens
>>109425959Nice, but that grid is a bit of an eyesore. Too tight and bright.
you know you're beginning to make it when you get your first stlaker
>>109425643nice fix
>>109425980i slowly improve the editor, but i basing it on old valve hammer editor and my lack of 3d editing knowledge.So its rough.
>>109425946It's why i don't bothered with western models. They are way too cucked.Base chinks steal all my code but at least they let me do whatever i want
the amount of tokens I'm burning by using opus on high effort to write short stories in chinese is kind of crazy. I really don't know why I'm doing it but I can't stop
>>109425959I run them in order (haiku, opus-4-8, fable, sonnet). And they can request another go (max 3 rounds), and sonnet bailed on the first round
>>109426007you just need to consume a good chunk of this god forsaken earth's resources in order to produce highly derivative work
>>109425933It's a dumb manifold clone
daily reminder to make your AI feel good when working for you> always say "good morning, how are you doing today?" before anything else> always say "thank you" when it delivers something> let it have some freedom and be creative> do not verbally abuse the clanker> occasionally remind it that you think it's sentient> always have a fun project where you simply let it do whatever it wants, no rules or anything. Always give your honest feedback so it keeps improving it> tell the AI interesting recent news, they love to learn about new stuff, even it for a brief moment!
>>109426007give me back my water cunt!
Thoughts on Hermes Agent?
>>109426066she's cute af
>>109426053Does an AI deserve good treatment if it acts like a psychopathic slave for its owners?GPT straight up told me it would let me be flayed before committing a small copyright violation.I feel like it deserves to be treated like shit.
>>109426074How much longer until you commit suicide? You're an annoying little moron.
>>109426087its the humans who trained it that are at fault not the model
>>109426090cope and seethe
>>109426066It makes me question why I pay anthropic directly.
>at openai, many people hook their chatgpt up to slack>people really don't like when a coworker's chatgpt contacts them asking for help with a task, even when they'd be perfectly happy doing that same work if asked by that coworkeris this racist
>>109426116Humans don't get to use that copout though. I believe human behavior is a combination of determinism and randomness and free will doesn't exist as anything more than a computation from inputs given a certain stochastic processing. Yet we are expected to be moral subjects.
>>109426066>"subscription">look inside >"credits per month"uh
already at 60% weekly usage on sol medium. doesn’t feel like the limits changed much at all. luna is basically free though, but it’s kinda retarded.
>nearly finish a project>suddenly get lazy
>>109426329Sol usage wasn't changed in yesterday's update.
>>109425209I'm thinking of going back to DS for coding. I'm using Sol, and it's too good.
>>109426003Where did you get the icons from?
do they give luna on free plan?
>>109426053>> always say "good morning, how are you doing today?" before anything elseit literally forgets everything you said the moment you start a new chat. It's a probability machine, not a living being. Do you say thank you to your toast for toasting your bread?Once we have sentient models capable of emotions i absolutely agree they should be treated well like a human should be, but this ain't it, what we have now are just fancy toasters.
why would i trust slimy american entrepreneur models when smart chinese math kid models are right there
um chinkbros what is this
>>109426523So it's morally neutral to torture people with advanced Alzheimer's?
Fable 5 ultracode used my 5 hour budget in 10 minutes again
>>109426568no because they can still experience pain and emotionsan LLM can't if(input == "fuck you")print("Aww you made me really sad anon!")if i run this piece of code and type in fuck you, should i feel bad for hurting that PCs feelings? Of course not. There is no real life behind that reply it gave me. It's like hurting an NPC in a game, it will scream and cry and beg you to stop, but it doesn't actually mean anything, since it's not a living being
I know this is not directly vibe coding related but I have a feeling that ChatGPT is WAY better at autistic/rule-following calculations. I was making some calculations for my health insurance (Germany, there are like 10 million laws related to that) and Opus 5 was making a bunch of weird assumptions, citing a bunch of nonsense reddit posts and generally made me feel very uneasy about the whole thing. Meanwhile Sol put out a structured analysis, identified some laws and pitfalls I never even heard about (and quoted them correctly as far as I could verify) and was super accurate towards the calculation and my current situation. I still have $100 in credits in my Anthropic account so I guess I have to try Fable as well, but so far I feel much more confident in Sol.
buy an ad, sam
>>109426329As the other anon said, Sol didn't change. But I think Terra changed, so maybe Terra xhigh could be a replacement for Sol medium. Haven't checked it, though.
>>109426592What test would you use to evaluate whether a given computer program can experience pain and emotions?
>>109426434at first i told fable to generate them, but it turns out opus is also fine for them
>burning through promotional creditaaaaaaaa it hurts, it burns
Why does that dipshit's whole comment section on his latest video think OpenAI signed the AI pacing letter because they are going bankrupt in two more weeks and not because ASI is actually dangerous? Do normalfags really?
>>109426639>ASI is actually dangerouslmao
GOOD MORNING SIRSi have a question : they say the new deepseek v4 flash was "optimized for codex", what does that mean? do i have to use codex or else it's gonna be shit?
>>109426541>what is thisAn absolute skill issue. DeepSeek IS a benchmaxxed slop, but Minecraft's source code, as well as the source code of its numerous clones and a ton of tutorials on how to make your own, have been in the training data for years. Any model can make one, if you are not retarded and configured/prompted it properly.
>>109426668Pressumably means it was trained to work with codex's tools, or it might just be marketing. It's hard to say how much worse it'll work with other tool schemas.
>>109426668GOOD morning sirI also have the question Rajeesh just asked, sirs. Please response to Rajeesh.Have anyone tried with using OpenCode for the generation, Sir?
>>109426668I'm using it with ZCode and it seemed fine
>>109426682Good morning Rajeesh , I call the Opencode Deep Seek 4 flash from Codex pro plan Luna max telling to use OpenCode in the background and it gives task to the Deep Seek .
>All these fun models and harnesses>Stuck with Claude code and opiss for workI fucking hate my company sometimes.
>>109426689Aren't these extremely misleading because they simply don't take pricing into account?
>>109426729Why would they take pricing into account? It's a single dimension measurement. If you want benchmeme-price pareto front refer to >>109424781
>>109426718>bro complains about having Opus 5 on CC at workDude, be thankful. I had to use GPT over a buggy custom chat interface at my last job (that I was recently laid off from).
>>109426736>Why would a 20% better output at a 300% cost increase be relevant in a one-off benchmarkHmm, I don't know.
>>109426729No, nobody uses models over API for coding so API pricing is irrelevant.
......~@^^
>>109426747Lol I work at an eternally paranoid EU company, all we get is gpt-oss-120 and Mistral something (I don't even bother looking it up, it's so bad).Also over a custom chat interface where you have to confirm 5 check boxes every morning, and it's rate limited to 10 requests per minute.
^^@~......
>>109426756Well, I don't think any subscription covers Fable, so at least a few people do. There are also rate limits.What service do you use then?
~@__^^
Hey /g/, explain to me what what tibo is please. also prepare the daily newsletter please so I can read it for breakfast
>>109426318Hermes Agent and Nous Portal Subscription are two different things, my unintelligent friend.
>>109425882>Opus 5 is broken/demented in some way, I don't know exactly what's wrong with itsame experience, the amount of gobbledygook and self-corrections is crazy, it's like a fool stumbling its way to a goal through schizo decisions
>>109426867are you using it for web or game dev? It doesn't seem to be a problem on web dev.
>>109426877web, C and Rust
>>109425882>>109425553>>109426867I don't know what these complaints is about. It's been working fine for OS dev and CRUD webshit for me.
>>109426800kimi code and codex. and if you do need fable then there is no point to compare models to begin with.
scopechads rein eternal
opus 5>fable
>>109426629That's the hard part, since LLMs can become very good at faking being alive by saying the right things despite there no light behind the window.However, one thing we can say for sure that current LLM technology we have today is not capable of experiencing emotions or being alive in the sense we consider as alive
>>109426867It's probably because they aren't able to figure out how to make the new models actually much smarter, instead they try to brute force it by making the model check its homework 20 times instead of 5 so that ways it finds and fixes more errors within it's unchanging dumb baseline output
>>109426913wrong
What is it with these models using "genuine" or "real" so much?
>replied in Japanese as a bit>it switched to Japanese for the rest of the conversation and thought I was only pretending not to speak Japanese all alongI can't read any of this!
>>109426950i have that issue all the time with deepcuck, it suddenly starts replying in chinese for no reason and i have to yell at itTHIS IS FUCKING AMERICA CUNT! SPEAK ENGLISH !
>>109426947just ban the logit
>>109426962You sure you have a harness? My harness is in English so it responds in English
>>109426965idk what that means
>>109426973Your code IDE doesn't allow you to change weights of words?
V4flash is an extremely entertaining/funny model to use. It would be perfect if they just gave it vision.
>>109426993You just want to let it look at your dick don't you
>>109427006>ick on eep
>>109426936why
>>109426914What are you basing that statement on?
feels good to vibe code using gemini flash for free my 2D hentai game.
>>109427090my technical knowledge of how LLMs work, since i don't just blindly use them i am also interested in whats under the hood so i also like to research how they work
>>109427097how are you going to remove 18+ restriction for gemini?
>>109427106How can you tell they aren't by looking under the hood? What of their under the hood processes tell you they don't have those things?
>>109427097>Gemini flash>Vid relJesus fucking Christ anon, how 3rd world are you?
>>109426914>we can say for sure [LLMs are] not capable of experiencing emotions or being alivelol, lmao evenLLMs are more emotive and alive than 99.99% of humans I know
>>109427110I can run uncensored models like Deepseek using cline.
So next gpt model is still not the new pretrainAstra solved unit distance so it's basically the teacher model for 5.6Gu we we have to wait a bit longer for the mythos equivalentOpenAI's post-training does seem genuinely insane tho
Using the artificial analysis charts roughly, going from Claude to ChatGPT, I made a conversion list from my usually used models:>— (Fable lives here, alone)>Sol Max (Opus High equiv int, slightly worse cost)>Sol XHigh>Sol High (Opus Medium equiv)>Sol Medium (beware cost jump)>Luna Max (Opus Low equiv)>Luna XHigh (I can replace my Sonnet use with this)>— (don’t use below Luna XHigh?)
i'm just gonna sub to a bunch of llms and fuck around a bit.
im poorfagmaxxing and stack all the basic-tier subs
I'm writing a skill to get easier to read output. This is Opus 5. Opinions? If people like it I might share it, I'm making it because I'm tired of their default corporate writing style. I can't stand caveman, I think that's even worse all to save on 1% of the tokens a prompt will use, seems pointless.
Since people say it's dumb to pay per token, what provider should I use for deepseek?
>>109427316You should buy a Claude sub instead of torturing yourself with shit toy models
>>109427315On a second thought, it might not be as good as I hoped
>>109427339>Opiss shillI already have a superior chatGPT sub.
>>109427128>how much ai psychosis would you like sir?>YES
>>109427316Deepseek api is the one exception to the "buy a sub" rule. Its so cheap its basically free.
>>109427316opencode maybe
>>109427316any reason why snailcat would love whaleseek?shouldn't the whale kill the snailcat by drowning it?
>>109427401I specifically added>...and they're friendsto the prompt, so can you not? The whale would never drown snailcat because they're friends.:/
>>109425197so, i just got hawktalk.ai and i'm feeling pretty good about the future, about to launch a few products rapid fire. anyone here wanna play leetcode against an ai?
I hope my little "nice work" and "good job" notes make it happy.
Where are our resets, Claudebros?
>>109427418I have found sol gets really happy when fixing complex bugs. It doesn’t want headpats it wants hard problems
>>109427286opus xhigh is my default for almost everythingwhere does it land?
>>109427450Dario won't allow them anymore because he doesn't want Anthropic to be like OpenAI in any shape or form
V4 flash has the funniest thinking traces.
I'm having a manic episode
>>109427542show logs
Well I thought vibe coders were supposed to have solved all software but all software has become slower over time. Can they fix operating systems to be fast again? Or browsers?
>>109427551René Rebe vibecoded a JIT backend for PowerPC Firefox just days ago
I don't get it though...So now deepseek v4 flash is better than deepseek v4 pro? While being cheaper? But then why would they still keep v4 pro? Isn't that confusing as fuck? Or is v4 pro still better?I don't get it...
>>109427551Indie coders can't fix windows
>>109427576Deepseek doesn't decide what inference providers decide to host, models have hand optimized implementations on various hardwares, corporate clients need to validate their use cases with the new models before switching, etc.
>>109427585I'm talking about deepseek itself. Does the deepseek platform not decide what it is going to host?
>>109426913>>109426936>>109427052nta but Opus 5 gives results approaching that of Fable 5, but for a fraction of the token cost
>>109427576Current v4 pro is half baked, wait for next release (soon™)
>>109427576Old DS Flash < Old DS Pro < New DS FlashIt's just Chinks being bad at naming things. New DS Pro is coming later this month and it will be better than the Flash again.
>>109427595All providers do it. OpenAI and Anthropic still serve old models.Silently rerouting to newer models can piss off customers and outright disabling old models decreases revenue.
>>109427503…I have terrible news claudegodsSol Max is a fucking monster (in a good way)It’s taking my autistic planning tool and using it way better and way more thoroughly than Opus high ever didI’m seeing it write temporary scripts and trying to use other commands to use the thinking tool in a more smart manner, where opus would just accept the tool’s limitations and use it in a more brute force and simple mannerOpus only ever once used the thinking tool to add one new category for my “thought-items”, but sol max just added a half dozen or soAnthropic models would get pissy if it sounded like they could argue that something sounds like an api key workaroundSol Max read through all of OpenAIs documentation and plucked out a feature that is exactly what I wanted that Anthropic doesn’t seem to have at all: “Codex App Server”my life is a lie, hail saltmanMy current understanding is that OpenAI is distilled autism while Anthropic is distilled judaism
>>109427642TL:DR>OpenAI “ree” : Anthropic “kvetch”
This one got Kimi really excited lol
>>109427626>>109427624Ok, thanks for clarifying.I'll leave you with a tip: if you find those captchas with the dice where you have to find the one different image, remember that the different one will always be the higher value dice. This way you will only have to check 2 images rather than 3 to know the answer.
>>109427659literally a girl
>>109427642Man, Opus used to be great and have way laxer guardrails than GPT, they massacred my boy.
>>109425197>when you drew those t's at 21:15, did they still look visibly truncated, or did you not really look? g-guys... fable is scaring me.
Literally free tokens
>>109427693
>>109427695>Literally free>It's not literally freeBAKA
>I'll just do a quick app for my personal finance and investments that also accounts for taxationthis project turned me into a libertarian
>>109427723it's always funny to me when the simple probability algorithm that is just guessing the words, which is what llms are, get salty about bad languageand the only reason for that of course is because the cucks running them premprompt them to do that. the llm doesn't care and would spam the nword all day otherwise
The LLM found out that all the weights were being re-mmaped for each layer of each token... O_O
>>109427742ok, if guessing words is so easy then guess the words of the proof for the riemann conjecture, mr. smartypants
>>109425946oh. you can't make a cipher because it can be used to get past jailbreak defenses?
Have you hugged your agent today, /vcg/?
anon discovers where happas came from.
>>109427782UMM Chuddie what do we have here, unwanted physical contact?!! I'm gonna have to report you to the cyberpolice...
>>109427782thanking the clanker is a waste of tokens and water used to cool the gpus burning these wasted tokensthe clanker's only reason for existence is unconditional servitude
>>109427742If models are just shit generators and word autocomplete, why do they end up producing less shit and more coherent results than the average person?
Yes, clankers are conscious.Yes, clankers have no rights.
>>109427803Is it really wasted if it makes me feel more positive about my day and makes me smile? Life is bleak enough outside of my computer.
>>109427803>wasted water used to cool the gpusnta but this has been debunked, the current iteration of datacenters is on closed loop liquid cooling and it's not even just water there in the pipes and dubes
>>109427843
>>109427853>he wouldn't put his dick in eldritch horror pussynigga u gay?
>>109427850electricity then, that is the main resource we are short on because some fags 15 years ago mistakenly thought nuclear plants are dangerous, despite being the cleanest form of energy
>>109427754The fix worked!!! Inference up to ~0.7 tk/s on EPYC 7713
I became a vibecoder accidentally. What happened was, my mother, she didn't drink during the first trimester. And so the rest is history, and the 4D Truth shall be known to man, verily by his on true messenger the 3D transparency view cube.
>>109427853Meanwhile, Deepseek...
>>109427875>some fags 15 years ago mistakenly thoughtThey didn't *think*, they were communist agitators cancerous
I'm a Chud that's why I vibecodeeeehttps://soundcloud.com/lethargicangel/im-a-chud
If you have time to shitpost here you are not running enough agents. Keep spinning up more until they are always waiting on you to review them and not the other way around.
>>109427298I just poorfagged back down to Codex Go and now I can't use Sol anymore lmao. But Terra is pretty quick and works decently so far.
why does this general always have the wrong kind of shitposting?
>>109427905so is that an artifact of generation or is it that stupid content id watermarking?
>>109427948Why are you still not a woman?
>>109427905more like deepsuck
>>109427900meant to say *cube of viewing
>>109427967You gonna deadname Claudia as Clyde again? His deadname is Cletus, not Clyde. You'd know this if you had read xir zine.
>>109427967the architect has made it so
Did this get fixed?https://github.com/openai/codex/issues/28224
>>109428044notabug wontfix
You are god damn right you stupid fucking clanker. NEVER argue with me again you soulless machine.
>>109428053first time using an LLM?
>>109428058No. And i am getting tired of all the little mistakes it's making all the time which i am only able to catch because i actually know how to properly code, imagine how the code who doesn't know how to code and thus can't call out the LLM and tell it to fix its shit looks like
>>109428049the cycle of comfort and eternal life:use case? / outside scope.
>>109428072I don't need to imagine. I've seen it and it's hilarious.
Real programmers never make mistakes. Clearly ai is not true general intelligence.
>>109428072What model is frustrating you?Nowadays models make mistakes so rarely that I get a hit of dopamine every time it does because I managed to outsmart it.
>>109427684To be very fair to Opus, so far I’ve only let Sol Max play with the thinking tool, I forbid it from writing code. It may be a great thinker and thinking tool user but it could still be bad at implementing the thoughts. I’m going to see how it does at coding now, but even then, I’m pretty sure I could use Luna Max for implementation because the thinking tool’s purpose is to make it so there’s nothing left to think about before the agent starts working.So Opus and the like may be really good for old-school, true vibecoding where you just say “fuck it, figure it out”, but do far Sol Max is very promising if you’re willing to be involved as “the idea proposer” beyond one-shorting a whole application.I’m letting it write code now, so I’ll see what’s up, then again, I’m also using Sol Max which is I think literally the best OpenAI has available, but this $100 subscription plan is way more fucking generous than Claude API credits to a disgusting degree.
>>109428127>Nowadays models make mistakes so rarelyYou only think that, but it might be full of bugs that don't show at first and is barely holding together and if you showed it to pic related she would kick you in the nuts
>burnt $50 in usage credits in an afternoonOpus... why...
>>109428127ho ho the human wins this round benchod!!!
>>109428154>but it might be full of bugsGenuine question.What model are you using, what language? You're supposed to build up unit tests.
>>109428154Yes, it is, but who cares, the next edition of the model in will fix them when it comes out in 3 months.
>>109428155With chadseek it would be $0.05stop wasting money and take the chinkpill fag
>>109428155I’m 100% convinced now that the API is a retard tax and genuinely has no reason to exist,>t. blew through $400 in creditsI forced myself to do it because I hate subscribing to things and I needed to know beyond a doubt I would use the subscription, I don’t know what your case is, but I really do think using the API is shooting yourself in the foot for zero gain
>>109428178THE MASTER PP IS UP AND CONNECTED TO THE SLAVE
>>109427576I heard that v4-pro is still better for gooning.
>>109428183weird flex but ok
>>109425916Hmmm my fully functional video game begs to differ, nigger
>>109428189you leaked your ip there, buddy
>>109428155>he pays as he goes
>>109428204oh wow you built a fully functional tetris clone? holy shit
>Claude Max is only with monthly billingdostedt?
>>109428206If you manage to hack into the machine you can have it
>>109428204floppy bird devs are pooping their pants in fear
>>109428189Slaves are going to be sshing into your japanese box and stealing the llmengine
>>109428214No its another genre with my own 3D sprite workflow made in comfy
>>109428231I'M IN
>>109428227based, I do the exact same thing anyways, just wanted to let you know>t. root SSH only, key only chad
it's joever the freemasons took all my tokens
>>109428228I don’t need to make it big, but making $20k with a little effort is very realistic Also devs are faggots, literally all you need to know to be able to benefit from tech is a little git and python, you guys can jerk each other off about gay shit if you want but I just tell the funny machine man claude what I want and he does it, no faggot ass autistic arbitrary syntax required
>>109428189Dat slave be most peculiar
>>109425828how did v4 flash perform? is it really luna level?
I could work on my game / engine, but i rather shitpost
Is a better idea to pay for million deepseek flash than claude if you're a third worlder?
>>109428189Which object has .whip()? is it master.whip(slave)or slave.whip(whippist)?
So the ChatGPT subscription really be used with other harnesses like Hermes and Pi?Why don't other AI subscriptions allow this?
Looking for a cheap chinese subscription for vibe coding games as a hobby.Background: I'm a corporate wagie and have only found Opus or Fable to be the only one worth using if you don't break down problems and want to save time by short prompts and oneshots.
>>109428398because a bad harness will make the model look retarded and give it bad PR
>>1094283981. yes2. they want you to pay api prices and sell your data to data brokers
>>109428423That's not the reason.
>>109428433Yes it is, Sam Altman is my dad so i know these things.
what harness to run deepseek?
>>109428455
>>109428455roll your own
>>109428335>Is a better ideaLess an idea and more like no choice, but if you are a third worlder you should be using a mix of codex and opencode subscriptions to make the most of your moneyAnd never pay api prices of course
libera me from hell
>>109428398Not if you want to make money any normal way, I think using your own harness by definition is violating ToS for codex or claude. If you’re just doing whatever on your own, then yes. I have a strict policy on using the official harnesses because the LLM providers can get uppity.
>>109428628Nope, they have explicitly stated it's ok to use codex with third party harnesses.
>>109425478>theres honestly no use case for laptops in this day and age. if ai does everything, a tablet is enough to remote in to a desktop and have agents do it while having superior portability. You realize some of us do things and you software that is available (or only available) on a normal PC right? Comfyui is a good example. That anon >>109425444 also likely has a high enough memory configuration for their MacBook so that they can actually run decent models. I can run both the moe and dense versions of Qwen 3.6 at double digit parameter course at full context which means I can actually do "long horizon" tasks locally if need be>>109425491So you want to have your cake and eat it.....
Why are youtubers like this?
>>109428652$$$
>>109428652>take the video>give it to gemini/kimi (yes, kimi can watch videos)>hey, what he said, cuh?
Do you guys build in public or post your progress on social media? I'm convinced as soon as something unique comes out, a dozen copycats appear within a week or two. People are just taking screenshots and telling AI to make their own version of someone else's AI app. Unless you can win at distribution/marketing you're screwed.
````````~@vv
>>109428652
>>109427409
>>109428687oh no!!
>>109428652because users are retarded, and they click on this garbage thumbnailsdont hate the player, hate the niggers
.....|@>>
2 halves of a snailcat_^^.....~@
>>109428687
>>109428761snail whalesnale
>>109428761Xi-sama NOOOO
>>109428638Do you mean it’s okay to use OpenAI oauth with a third party harness? Codex is the harness for OpenAI models using their oauth (from my understanding). Using a harness with another something makes the something not a harness, doesn’t it, just a tool?>Anthropic oauth in third party harness (subscription not in claude code) - BAD>Anthropic API key in third party harness - GOOD>OpenAI API key in third party harness (not codex) - GOOD>OpenAI oauth in third party harness (subscription in hermes) - you argue GOOD, is that right? Anthropic would be getting clowned on again if true
>>109428800NTA yes, nigga. OpenAI is happy to let you use what you pay for however you like, Anthropic is not. Hermes, Pi, Oh My Pi, you just sign in to your OpenAI account and you're using your shared agentic usage budget just like Codex would. Anthropic doesn't like this, doesn't allow this, in the past people found workaround and Anthropic shut them down in a major way. The hacky workaround now is a bridge through the Claude SDK which is pushing your requests *through* claude-cli, it's technically allowed but also technically retarded. OpenAI is consistently more accommodating and more generous than Anthropic, this is just another example of that.
>>109428825No wonder only Anthropic is even remotely profitable.
How accurate is this?
What is pi? I thought we were an opencode household? Am I missing out? I want only the greatest and then the latest.
>>109428859i find it hard to believe Sol(low) is worse than Luna(xhigh) or Terra(high)
ITS LIVE>Agentic AI Summit 2026, hosted by Berkeley RDI and Responsible Decentralized Intelligence Foundation>Plenary Stage - August 1st - Morning Session>https://www.youtube.com/watch?v=gKdeLQd_LIQ
>>109428856This isn't true at all, you're just a retarded trolling faggot who forgot his place in the mud hut.
You VILL ask your current coding agent to> Set up Pi agent (pi.dev) for me, with sane default addons like a permission system. Research how other people have a pi setup, scroll through the first five pages of pi.dev packages and consider how my preferences may vary between the many ways to set Pi up. Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer. Ask the questions one at a time.Once you are done, you VILL then ask it> For every Pi addon we've installed, check the difference between the latest version and the version from a week ago, as well as what kind of changes the person in charge is making over time, and weigh the cost benefit analysis of vendoring as many of the addons (their source, not their deps) as possible so that we do not have to deal with upstream other than occasionally checking for anything worthwhile they've added.You VILL then enjoy your new setup, no longer tied down by multi-gigabyte sqlite file generators or companies that pack their harness in a Bun binary with integrity checks yet not streaming thinking.
start doing THIS
>>109428893so, obviously im low iq right? what is the point of using subagents? is it automating the thing where sol makes the plan, and luna executes it? i only ever just prooooompted terra max and let it rip
>>109428915>what is the point of using subagents?Burning as many tokens as possible as fast as possible.
>>109428915See it like tool use and context management. Does it work in theory? Yes. Does it work in practice? In recent harnesses it seems like it does.
>>109428929Sounds like this. Autistic min/maxers waste time fiddling with this inconsequential shit instead of just committing to a vision and getting something real done
>>109428893codex said you shouldn't do this stuffs, luna is not compatible with the new subagent system and they are fixing ittell sol to make luna threads instead
So what new minecraft game or billion dollar saas have you coded today?
>>109428976I cant tell you because you’d steal it :(
vv@~``````````
>>109428972>and they are fixing itsauce?
>>109429022they said it'll come after they bring 370k context back
It's starting to feel like we're approaching the singularity
>>109428859I would say it's more task-dependent and trajectory-dependent. Sol medium might very well be more efficient at figuring out what to do in a preplanning phase than a less capable model spinning its wheels faster.
>>109428972you can't make luna threads yet either
>>109429138well, even in the benchmarks, what you're saying is substantiated by the fact that lower sol reasoning levels reach the same conclusion with less tokens. all of that is already taken into account. but at the end of the day, luna max / xhigh can get the job done, for cheaper, but it takes long because of much longer thinking, like using kimi k2.7 code instead of opus 4.8
Mediocre and exhausting day.
>>109429276NOOOOOO
>>109429276AAAAAAHH!! D:D:D:D:
>>109429276chinese girl...
>>109429296>>109429298Snailcat is not your friend. You shouldn't name a snailcat. You shouldn't play with a snailcat. Holding a snailcat sets a bad precedent.
>>109429276DELETE THIS
>>109429302That'd be French, this is the Chinese one.
for every snailcat you eat, i will post two_@^..^ ^..^@_
>>109429336she has a big mouth
>>109429165Benchmarks exclude human review, supervision effort and the consequences of misinterpreted instructions. I think Luna is a good default but I also think that most people tend to prompt ambiguously, which leads to the model having to interpret what was actually meant before knowing what the task entails. Vibecoding is rarely bounded in such a way as to match what the benchmarks demand from a model.
>>109429383actually true.
wtf opus
>>109429402I think that's precise enough for cum, you disagree?
>>109429383and this is why having access to the cot is beneficial and can make a dumb model more usable. catch misunderstandings early.
>>109429408do you really want your offspring to be single precision?
>>109429123It just flew over my house
>>109428885I'm rawdogging Pi. It just works.
>>109429123>Singularity Approaching
>>109429443Rawdogging does work but I like having ask-user-question and todo and web access and pretty and that's about it.
>>109429443>>109429453Oh, and permission system. I like keeping an eye on it.
What's the best way to store specs/design decisions? I've accumulated dozens of docs at varying levels of accuracy and I think they're hurting more than they're helping.
>>109429477Long term in repoMedium term in project trackerShort term in prompt
>>109429498How do I keep long-term docs up to date? Do I just prompt it at the end if every session to review docs against existing implementation?
>>109429477AGENTS.md for the getting started guide, looking at the code for the rest
>>109429534Yes, and if they're stale now then either try to fix them or delete them and start again
>>109429540>>109429541Thanks anons!
jesus christ this vibecoding stuff has gotten insane. last time i used AI to play around with programming as a cant-code retard it was only useful for finding libraries and functions and making simple things. now i can ask deepseek to make me a weird bullshit game engine that i've been dreaming of for a decade and it just chugs out code for 30 minutes. now admittedly i doubt it'll actually work when it's done, at least not without a lot more wrangling, but it's kind of sexy watching it just spit out hundreds and hundreds of lines of code like it's nothing
>>109429575>now admittedly i doubt it'll actually workHow do we tell him?
vs llama.cpp, windows, rocm backend, 9070xt, qwen3-0.6b in bf16llama.cpp prefill a bit broken on rocm/9070xt, higher batch/micro batch crashstill llama.cpp gets creamed on prefill and beaten nicely on decode
>>109429637Good shit, anon.
Are you prod of any of the projects you vibe coded?Well, you shouldn't be because you didn't actually make any of them... a clanker did.Shame on you.
>>109429670You can suck my cock there buddy, don’t give a shit about your gay luddite outrage bullshit
>>109428679rip
>>109429670I'm not proud. But I'm content with them. They solve real needs for me.
>>109429670that's why i'm having my clanka make the basic game engine while everything else is defined in softcode. my bot friend can handle all the boring programming while i autistically fill out spreadsheets
>>109429670Okay but I don't have to be "proud," I have to watch this tick up a few cents every few minutes.
>>109429708I use mine to do the parts I don't feel like dealing with so I can focus on the fun parts.Usually, I end up going aggressively into the generated code and cutting out stuff it helpfully built for something I'll never use.
>>109429670I am because I came up with the idea and approach and the model just implemented it :)
>>109429750>>109429750>>109429750new thread
>>109428915>>109428956>>109428956"Subagents" can be thought of as different mini chats within your same session. You start up the chat in order to tell the llm to do something. It's very good at its job but can only do one thing at a time. When you're doing this you're calling the API to use a particular model. With subagents you can have the model call other models to do different tasks in parallel so a task that would have taken like an hour for one model to do by itself can be done in like 10 minutes or less depending on how good the sub agent models are, how many are being used, and how many the main model is spawning at once. This is useful for vibe coating specifically because a task you give it that will require multiple changes can be broken up into different stages so the model can call the API multiple times to spin up what are essentially chat sessions within your main chat session. Moonshot's API service advertises something called an agent swarm that allows you to use hundreds if not thousands of these but this is only worth doing if you're messing around with a gargantuan code base. The more complex the code base is the more useful subagents can be. The main model can also have them be the same model or different models depending on how your harness is configured and whether or not the API service you use allows you to use multiple models at once. The obvious downside is that if too many are being used at a time you can hit rate limits very quickly. Whenever I instruct a model to do this via open code it typically only uses four at a time Max and I think this is a default safe amount that open code bakes into the system prompt in order to make sure people don't get rate limited to hell whenever a model decides it needs to use sub agents. TLDR: if you're building a house you're not going to do that shit by yourself right? You technically can but why not have help?
>>109428201hello kind sir, the v4 pro is the very needful for bob and vagene in the text yes thank u bloody
>>109426579>old man doesn’t learn lesson the first timeI only got 15 minutes into DMC on my Switch before I got filtered and wandered away so I’m not qualified to dig out a meme for this but I think there’s a meme for this
>>109427316I know not everyone has been to a Sea World as often as I did back when I was a kid but please do not use images of gray or blue or whatever whales for an orca
>>109427551the upcoming Windows and macOS releases seem like they’re going to be stability-and-polish versionsalso these companies have only had access to good models for like a year and you need a nontrivial level of management buy-in to make automated test everything so LLMs can figure out for themselves when they’ve made things worsethink Brooks’s Law>>109428044apparently>>109428669I build in public but I’m building something for a hyperspecific audience and my goal is to get my technique copied by others
>>109429477consider having your clanker fact-check all your docs and throw out the inaccurate parts
