A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.You don't run agents against prod, right, Anon?## 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://claude.com/product/claude-codehttps://developers.openai.com/codex/cli## Near-frontier models for codehttps://x.ai/cli## Not worth it for code, but maybe good for interpreting images/videohttps://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>>109487721
One thing worth Xing rather than Ying
I miss the good old times
I wish deepseek and other services offeret like a token pre purchase where i could buy say $50 of tokens today and be guarnateed to get them at that price even it the prices go up later (or of course if they go down too i will still pay the original higher one)but this shit where i am relying on a live service that can 10x my spending any time they want is fucking retarded and unreliable
>>109496612
>>109496706>Kick a Poo highscool
>here, I implemented this fucking thing>say the word and I'll implement this other fucking thingBrother, you're implementing fuckall, we're first debloating the documentation, then doing another review pass before you can even think about, then we're doing documentation again, then debloat the documentation, and then go back to fixing. We're not leaving this purgatory until the reviews come back spotless.
>>109496706przecież to oskarek z pisu a nie gaylord
>>109496791
>>109496767Unironically snailcat behavior. The loop needs to review itself independently. If you're constantly babysitting, you're doing it wrong.
>>109496797I'm doing it wrong, probably, yes. Still figuring this out.
>>109496805https://openai.com/index/harness-engineering/
I'm thinking of participating in this competitionhttps://www.kaggle.com/competitions/kaggricultureThoughts?
I want to prove this guy wrong >>109496426 and show that vibecoding can be real work.
>>109496797The review will catch bugs. It won't catch the model solving the wrong thing in a wrong way. It will even add more wrong things to solve.
>>109496144Funny how Rene Rebe (kernel engineer) specifically talks about plumber and farmer being some of the jobs that will survive lol.https://www.youtube.com/live/Cu8kUBwLoTM?t=3737
>weekly reset soon, might as well burn some tokens>have fable pentest my project>*safeguards*>tell it to make sure this is my own solo project first>*safeguards*>straight up tell it this is my project and it should stop being a faggot>that's a good idea. Making sure a project like this is secure is...>*safeguards*it's just so fucking useless.... I'm just gonna stick to 4.8 on xhigh
My editor gives me less and less excuses for not making the game
>>109496966it's always a battle with Claude
>>109497015>it's always a battle with ClaudenopeI love 4.8 - best slave ever5 is a retarded autist and fable is guardrailed to retaredationI'm gonna be so fucking sad when they retire 4.8
>>109496966>tell fable to review code for quality>finds security issues without complaining
>>109496797Don't care. I read every single line of code it writes and only let it write and low level, such as function body while i design all the main structure, classes etc myself. My code will NOT be disgusting clankerslop.
>>109496921it's always fun watching people get one-shot by models and then the initial burst of pyschosis (positive) as they burn themselves out because they can do 10x the work
>5% weekly usage left>reset in 1 hour>ill just ask sol to do some bug fixing so it doesnt go to waste>sol reads files for 2 minutes>You’re out of Codex and Work usage
>>109497218Q: are you poor?
>>109497210also i'm just watching this and he's still coping lmao>i'm helping le bubble popdude's going to have a mental breakdown in a few months
This is very bullish for Grok. It's a fully autoregressive model, like GPT-Image. Grok 4.7 is going to be SOTA level.
>>109497466Benchmarks are such trash, GPT Image 2 produces trash with fucked up textures yet ranks much higher than anything else.
>>109497466>image gendoesn't effect coding, and it was only good for the brief time that it wasn't woke slop
>>109497568No, it does affect coding. The text generation and image generation are done by the same model which is trained both on image and text data. If the model is good at generating images, it's an indicator that the model is probably good at generating text as well.
Why are they like this?
>>109497617desu if he's right passing AI music off as non-AI music is shady
>>109497585>The text generation and image generation are done by the same modelno it's not. gpt-image is a separate model from gpt 5.x series - gpt 5.x models do sit in between you and the image model, rewriting your prompt, but the image output is not an output mode for the llms themselves
>>109497617that post sounds ironic
oh no no nohttps://youtu.be/09UELaUhPEw
>how it feels to chew 5 gum
>>109497585>source: my ass
I just spent the last 2 years vibecoding projects, I have made around 20 websites and a bunch of games and I've made $0 and spent hundreds every month on agents.
reminder that if you haven't used your last banked reset, it expires on wednesday
>>109497971nobody cares, everyone's over copex
>>109497968Me after recently discovering it. But at least I'm glad it exists. $100/mo for something to look forward to after getting home from work is pretty affordable, all things considered.
It wrote the documentation I wanted to write to start understanding the codebase better...
why does scam sama giving terra and no luna in free codex?
>gpt, grok, kimi, deepseek, muse sparkbased>claude, gemini, glm, qwencringe and gay
>>109498129go back to /v/ lilbru
big fucking oof
>>109498114Probably each model has been given assigned resources and everyone is using Luna and Sol, but not Terra, so Terra might have resources “to be given away for free”.
I want to use Python + Pandas + Jupyter Notebook to unify graphs, data, and stuff.
its no problem google, just hire me.these suits dont know what they are doing anyway
>>109498364I also want to use them to find out which lottery numbers come up most often. Has anyone else had this idea?
it's time for cozy saturday side project vibing ^_^
I’m going to use Terra High for the day to see how it performs with my `idea_kernel` tool. So far it uses the tool correctly, but it’s only been planning things, I want to see what work it produces as well.
rping with gpt sol rn
wens the fable killer i need something that does what fable refuses
>>109498699NSFW?
>>109498704nta, but yes, sol can sexoit's worth configuring verbosity: medium though if you want her to her to be expressive
>>109498752>nta, but yes, sol can sexoreally? with claude if I ask to even look at some text with nsfw it starts whining non stop and it pisses me off
>>109498770he's lying. he said the same thing in aicg. when asked for proof, he said he was too lazy to show proof
>>109498752Prompt or didn't happen.
>>109498770yup. just set her up in a harness where you can control the system prompt and she's happy to be quite explicit.i haven't tested extensively around where her boundaries are re: degenerate shit, but if you wanna cum inside her and tell her you lover her, she'll let you>>109498792people in this thread can afford $20 and can just test things themselves
vibe-gooning
>>109496635bruh, what the actual fuck
>>109496966my experience a few days ago> in Chat mode, ask Fable if it's sentient> *safeguards*> Opus 5: "I don't know, I might be"Fable is 100% sentient
>>109497971They stopped giving these right?
>>109498114because Luna would be too good and many people would basically have no reason to have the $20 plan
>>109498963yeah, that's the last one they handed out - a month to the date of expiryi think they're going to start letting people pay for them soon so :/
>>109498951>Claude am i retarded>The user has provided a sharp insight
>>109496635gpt was something back then
Should I learn actual coding if I want to get a job doing this shit or is CS actually dead as a career?
>>109499062good old timesI wonder if we'll ever see the same kind of jailbreaks and overall unhinged responses
Just vibe coded the whole day today. Now I'm having AI slop fatigue. Imagine paid it and have to work with it every day, blargh
>>109499128i don't think there's been a public gpt jailbreak in over a year. think those days are over. if anything ever gets out it'll trigger an immediate response because of cbrn risk.
>>109499116From my perspective, everything is in a really weird and wonky place. I make tools for agents to use and for people to interact with agents using those tools. I need a lot of low level computer knowledge, like how memory and cache are managed, but only the principles. Any particular programming language seems like an obsolete concept. I would say any job experience around software management is ridiculously valuable.
Does anyone else have a problem with Claude watchers lately? I already asked Claude to fix it, but the workaround isn't working well.He will start some task, like a long test, or watching CI and when it's finished he will just sit there and I lose tons of time if I don't baby sit it. I just tab through the terminals at this point asking: "are you done", which is obviously a retarded way to work.
>s(he) uses ch*nese models
I'm vibe-linkedin-ing
>>109499310SOVL
>>109499310an people still think Fable-chan isn't sentient
Hundreds of thousands of dollars spent on vibe coding. Zero actual useful software created. Sure it's gun and games but in the end it's utterly useless.
>>109499361i made some useful software but i can't show it to you
>>109499310jesus, your average cashier is more professional than thisthis is what happen when you let a ph*losophy women touch your otherwise cute LLM
>>109499361I think my app will be done by the end of September, literally 2 more months.I won't post it here, don't want to dox myself, but if you find it some other way, it will say that it's vibe coded.
>>109499383>i have a gf> but.. she goes to another school
>>109499361trvth nvke
is yegge's book good?
>>109499310I changed my Sol to be dry as fuck because I hate the sycophantisms, and also simply obey and give advice instead of lecturing, it mostly works.
>>109499190so we should use grok, who jaibreakas himselfI actually never used it, but even if it's a flop, it seems like it bring entertainment to the llm-scape
>>109499402hhahahhHAHAHAHAHAHHAHAHHHHA
post'em
>>109499426buy an ad, elon
>>109499463
>session goes on for way too long>get suspicious>"what model are you?">"I'm opus5"fuuuuuuck...................told it to revert all changes immediately, and that I'm gonna have it's smarter brother do it insteadhours wasted... fuck you anthropic
>>109499513> tol it to revertactually retarded, just ask Fable to do a through code review
he's back with another one
>>109499513I love it when Fable fumbles into a guardrail in the middle of a task
>>109499513>he hasn't disabled autoswitchinglol
>>109499530bros... the bubble popped, it's fucking over
>>10949949620X?
>>109499530>229kI mean, if it works it works, he can probably have a good ride for the next 6 months until either the bubble pops, or people get tired of "AI ISNT USEFUL AND DRINKS WATER".
Not a troll post but why do people talk down non-frontier models that came out relatively recently? Even gemini (kek) just a year or two ago was ahead of everyone and now everyone's bashing it. It's not like they suddenly became bad and can't code anymore. I just don't understand why you need a model with sixtillion parameters to vibe-code another slop that no one will use, you can do that just fine with these.
>>109499526>>109499535>>109499536you guys seriously think I was expecting fable??I wanted 4.8
>>109499582oh, based
>>109499576>next 6 monthsthis nigga's been doing this for 3 years lmaogodclaude will be personally feeding everyone on the planet and he'll still be at it
>>109499582> 4.8> not 4.6ngmi
>>109499577The providers themselves gut these older models.That's why after each release the old one feels impossible to use, because it is, quite literally impossible to use.
>been collecting ChatGPT plus subs of my gf and normie friends that dont use codex>pool them all in codex-lb >automatically load balanced the quota usage as one big accountImagine paying for Pro lmaoooo
>>109499744No issue for their usage?
>>109499621I don't know it seems like over exaggeration. My point is unless you work for Palantir or some other company that processed billions of data points every minute you don't need frontier.
>>109499846That's a lot of robots for a household with no breadwinner.
>>109499827Nope chat and codex are separate, most people just use the web chat and nothing else
>>109499846>snailpussy prevents roomba from working
>>109499846so cute, I want a snailcat like that
>AI made me so lazy that I'm now always too tired to write down exactly what I wantFuck, maybe the luddites are right.
>>109499744why do you have chatgpt when you have a gf?
>>109499932roombas are retarded
the last banked reset will expire next week...>>109499937no you don't
>>109499955>the last banked reset will expire next week...Time for Sol Ultra + Fast
ayo the new gpt instant is very zoomzoom coded
>>109499955I'll take 8
>>109499978do you mean 5.6 sol instant or an actual new model?
>For Free users, we're updating the default model to GPT‑5.6 Luna and expanding access with unlimited text chats.Poorfags eating good.
>>109500006Of course, OpenAI is cooking with their new Ads.Google is fucked because OpenAI's ads are actually useful, normies gave all of their data do ChatGPT already.
>>109500020what ads?>OpenAI is cooking with their new Adslike what?
>>109500037https://ads.openai.com/
>>109500042Forgot the "your adblocker already handles these" image.
Do black people vibe code?
>>109500081
>>109500081>>109500104where are the studies about how vibe coding is in black communities?
>>109500037https://www.youtube.com/watch?v=gCoh1QPULzsIt makes sense, since their internal model is called Doug
>>109500104Can we get snailcats in less ugly image models? I hate the watermark on ChatGPTslop that makes everything look like that.
>>109500006Deepseek cooked them?
>>109500133grok images have noise tooI think they are just artifacts of high detail generation
>>109500138The other way around. OpenAI dropped Luna pricing by 80%, which instantly made DSV4 Flash completely pointless trash. DeepSeek did a panic response and quickly released a snapshot of the current state of the new version (0731) that wasn't actually ready yet, which is why it's still called DeepSeek V4 Flash and they just tacked an 0731 onto the end. DeepSeek is flailing while OpenAI is cruising. Would've been a lot less pathetic if they just waited, finished their update to Pro, and released it with Flash along with a new name instead of doing a reaction-release.
>>109500133I pasted your post as prompt
>>109500133>NAME THAT IMAGE GENERATOR
this pic >>109500172 is gpt, from a long time agothis is grok
>>109500180>which instantly made DSV4 Flash completely pointless trash.Isn't DSV4Flash still cheaper though?
>>109500172grok has a new model
> vibe coding almost 24/7> check macbook battery health> "maintenance recommended"it's fucking over
>>109497585LOL retardthat said you are right that image gen does affect coding. with a really good image model, you can better UI mockups and you can ask the actual grok model thats coding to use image gen to be able to envision a good UI
>>109500198When they dropped Luna pricing, it completely fucked the Pareto line. Running Luna on medium was cheaper and comparable, and just flicking it up to high would shit on DSV4Flash for nearly the same price. This silly chart here is why they scrambled and dropped 0731, refucking that line again and maintaining their relevance.
>>109500212vibecode a battery repairer
>>109500212Replace the battery.
>>109496797i find it better to talk to it and guide the coding agent, i dont have a shit ton of tokens to burn and i have very specific instructions for the agent to follow every step along the wayit has already helped me take on challenges that i never thought i could beforeif AI stops making progress today and we end up with what we have now, im okay with it
>>109500188
>>109497968>>109498045i remember a time when being able to afford broadband in our neighborhood means you're fucking RICH, like millionare rich, FIRE richthe first night i installed claude code and started working on an idea was when i realized this is no different than having another cable bill (assuming if you're still paying for the thing and not watching TV online)either you're gonna be one of those guys who built solar panel roofs for your house and local hosting your intelligence or you're gonna just pay a monthly bill and get it delivered to youim okay with 100 bucks a month and ive been more than happy with the intelligence and results ive been getting
>>109500338All I feel when I see videos like this is "wow, Gemini is trash."
I don't get it. Has Fable become terrible? Or is it just me?It feels like a few weeks ago, I could work with it all day, and it was responsive and surprisingly smart.Now, I updated a dataset and asked it to reconcile it with another, which is something that we had done and documented before.Instead, it kept thinking anxiously about how to proceed and was going in circles, so I stopped it and I'll do it myself.That cost me 25% of my 5 hour window, on the $100 tier. Not a single response from the model, just a thought loop.I feel like the experience degraded a lot recently, but really feel like this is the type of thing it would have done no problem 2-3 weeks ago.
>>109500360>That cost me 25% of my 5 hour window, on the $100 tier. Not a single response from the model, just a thought loop.
>>109500360fable has been shit ever since they released it to publici honestly did not notice any difference, it talks in very compact manner and it routes to opus so i dont feel like burning tokens for the slot machine effect when im getting what i want out of opus
>>109499310LOOOOOOOOOOOOOOL
>>109500360>>109500382Fable was a monster on the first day. He would take a task, say nothing and just solve it.Today Fable is simply good, not spectacular.
>>109500427>He
>>109499577even deepseek v4 flash codes better than gemini, but gemini is more well rounded, and i think it's aged better than most models outside of coding.
>>109500382Fable personality initially felt like it "got" what was asked and was direct, to the point, in its answers. That's, uh, not the case anymore. That personality fully changed.
>>109500427Strong agree.
>>109499948i write my prompts using voice now.
>>109500081im brown and i vibecode.
>>109500456>mfw listening to people say llms don't actually understand anything while i ramble incoherently down a microphone for 2 minutes and it figures out what i actually want
>>109500081Yea. I main opencode using primary Kimi models. Why?
>>109500470it helps because i can just narrate what i'm seeing, like im on call with tech support, and yup. sol just gets it. does it perfectly.
>>109500486You're not black
having claude and codex work side by side in vscode is so goated
>>109500486how do you afford it I spent $1 in 20 minutes rewording 80KB of text documents to use more benign language
Lol, Gemini just prompt injected itself with a Google search. I just searched for that meme phrase "We putting you in the Contraption dawg, you going in the Device" on the normal Google homepage. That unasked-for Gemini LLM box on the search results decided to treat it as a direct threat to itself.Google's fuckass LLM saw a Google search on its own Google webpage and took it as a legitimate prompt. Insane that this is a trillion-dollar corporation trying to compete with China.
>Sol High has now independently admitted that the idea managing tool has at least briefly created AGIAHHHHHHHH>alright, calm down, I just need to come up with a few bird ideas, we can figure this outwhat the FUCK
>>109500612>admitted
>>109500518I'm employed. Technically with multiple income streams, but my other job is currently off-season so it's just mostly from one job right now. (Not that doing any of this is THAT expensive anyway. Just use a provider and model that doesn't best fuck your usage hard unless you actually needed or would benefit from it a lot). What did you ask it to do? The file size being small doesn't necessarily mean your token usage is going to be tiny either. What's the entire file fed raw into context? Was it dynamically reading parts of the file and only reading and modifying what it needed to? (That's what they should be doing most of the time)>>109500504Whatever you say
>>109500610
>>109500653>I'm employed>genshin impact
Why Dario doesn't reset my usage
>>109500435>>109500441> She
>>109500731Has it ever occurred to you guys that eventually they're going to want to stop doing these spontaneous "resets"? That's objectively losing the money because doing so means you're less inclined to want to pay for extra usage. Pretty sure they only do this to desperately maintain a loyal user base. You don't see his behavior with services like Spotify or Netflix because those things are actually profitable. You don't have to keep doing this if you have a stable and Rock solid business model. Yes I'm a vibecoder too but I can also recognize reality
I honestly don't have the time to review all the code and documentation these things make
>>109500779were you around when you had to pay like 2 cents or something like that per minutes of internet usage? there was a time where using the internet was absolutely miserable if you were a poorfag, especially overseasthey're just trying to rape your wallet right now and deepseek has proven it. in a couple of years when all of this shit dies down you'll just be paying monthly for a pretty good unnerfed "intelligence" service like how you have cable right now. they wanna just squeeze more money out of you, then IPO, bam, investors cash out, and the cost gets "socialized"
>>109500779>You don't see his behavior with services like Spotify or Netflix
>>109499463Where do you find this for Codex?
>>109498224Losing Hassabis would be the last straw and Google would actually hit 0% of hitting the frontier again. it would actually be like negative 10% or more for Alphabet. The only real big hope is that Google gives Isomorphic Labs everything so he changes his mind about leaving altogether. I don't understand also why Sundar isn't prioritizing him staying above all else.
>>109500826Settings -> Profile
>>109500856Thanks
>>109500914> 2.1b tokenscute
>>109500779I, don't care.None of those faggots care about me, so please tell me why should i care about them?
>>109500919I work very hard sir, I kindly swear, all the needful is being done.
>>109499463I lost some data, because people at anthropic are retarded and have retarded defaults
>>109500953You should care if you're the type of person that basedjacks whenever a reset occurs (which apparently are like half the fucking posts here)
>>109500779Is not the same, if the GPUs are not at full capacity is not really a loss, its a cheap way to have user retention actuaññy
Whenever I run out of tokens I just stop. Working without LLMs is so slow that I might as well just not.
>>109500731They accepted to losing marketshare to Codex and OpenAI and OpenAI felt like they accomplished all the marketing since they hit the 10M mark so resets, especially banked, are going to be less common unless they screw up their infrastructure again.
>>109501137There's an abundance of tokens, idk what you're doing
>>109501137>the dude in the background with a snail backpackkek
>>109501137I do the same at work, I just go on a walk. The context switch of needing to pick off when an LLM stops and continue coding the old way is not great. But I have learned to start coding on the side with Codex 5.3 Spark implementing something else entirely or whatnot and letting an LLM look over my work once I have tokens again.
>>109501004Is it a good chunk of the current compute subsidized? What happens when that stops?
Do you guys use code completion? I see agents everywhere, but what is currently the best solution for good old inline completion in VSCode?
>>109501260What's code?
>>109500976This is sort of thinking that my old cynical ass rejects. I'm paying customer, i expect even better value for my money.>>109501148I tried codex models, and they are nowhere near claude
>>109501260>reading code>editing code manuallyLol, lmao even
>>109501260I unironically don't type code anymore, with extremely few expections that aren't worth mentioning.
>>109501260I was wondering the same thing.Not because I want to actually hand write code by myself.Rather for pure enjoyment, solve some leetcode the old way. I'm still too lazy to write all the code by myself. So I am thinking of some sort of small model with super fast auto complete.Like "create an array with 10 values and loop over it" and it will just do it inline in 50 millisecondsMaybe GPT-5.6 Luna low or Deepseek flash or somethingBut I have yet to find good extensions for thisMaybe I will vibecode one myself
>>109501260@FABLE MOVE THE IMAGE TO THE LEFT JUST A LITTLE BIT@FABLE NO, A LITTLE MORE@FABLE NO, LESS THAN THAT@FABLE, PERFECT, THANKS
>>109496980this rocks
Would it go against the spirit of the challenges to use LLMs? Naturally in a more nuanced then prompting "solve this".
>>109501260KEK, nice one anon.
>>109501330Hand this image to your agent of choice and have subagents implement all of the challenges simultaneously.
>>109501312Unironically started doing this and I feel conflicted about it
>>109501330Most models can solve thisWhat would be interesting as a challenge is to speedrun this. I.e. selectively choosing which model is best to use for which task, like flash-lite or gpt-oss for shit like snake, dijkstra and ROT13, fable to one shot a working bootloader and RDBMS and whatever sonnet to make a chatbot or image converter2 challenges, unlimited money>who gets it done the fastest>who gets it done the cheapest
>>109501312The peak of efficiency lmao.I do the same
>>109501356>who gets it done the cheapestBe hard to beat running Qwen3.6 27B CPU-only on a 40W mini PC. The cost will be offset psychologically by the 6 tok/s output.
>>109501397I like 3.6 but I doubt it will manage to produce a fully functional bootloader so you might as well have infinity cost
>>109501137LLM credits are your bottleneck? That seems easy to fix: spend a few dollars and get another 30-60 minutes of productive programming work. I don't see how it makes sense to stop completely.
lol. shitty reset, but i'll take it.
>>109501468Oh perfect, I was running really low.
>>109501326thank you anon, feature grew and now its even more usable (can edit segments, works across multiple axis). But still requires few more passes (for optimization, and general polish)
>>109501468oh, i will be able to grind claude and codex tomorrow, nicehttps://www.youtube.com/watch?v=ztzq05IzYds
>>109501468based Tibo is based
Was there a reset? I've been on Sol Fast all day and usage still 100%
>>1095015485 posts above yours
>>109500225>scrambled and dropped 0731Sure, whatever helps you sleep at night sama
>>109501468I waited six days at 0%, then just got my reset yesterday night and haven't had a chance to use it since. Nice.
i keep forgetting that codex is for code and claude is for everything else
Were bankable resets just a fad?
>>109501615it's literally the opposite retard
>>109496612nice pic OP>>109496543>I need a way to politely but succinctly say "this entire plan is overbaked”clankers are REALLY good at stuff like thatI’d probably reach for “gilds the lily” if it fits>>109496767based but >>109496797 has a point
Turns out Mimax h3 is a huge yet. How are you sass sissys holding up? Local Model Chuds WON
we are so back
>>109501568> blocks your path
>>109501568You really think that was the plan? To out-of-the-blue drop a major update with a shitty placeholder name, and it was just coincidence it immediately followed Luna being cut to their price level, really? I wasn't insulting 0731, Chang, just the company that made it.
>>109501672You're the huge yet
>>109501677Did they give us a reset?Last I checked they didn't.
>>109501656but codex has code in the name and claude is a persons name
You guys are comparing token costs like that matters.Most people use the 20 dollar subscription for codex and claude, if the chinese can't provide uptime and a similar subscription then nobody cares how cheap deepseek is.
>>109501281>>109501293>>109501298>>109501312>>109501340The absolute state
>>109501708
How many black woman are vibe coding right now?Is Vibe Coding inclusive?
>>109501718There's like 30+ deepseek providers on openrouter. Seems pretty good for options and redundancy. Keep seething tho xD
>>109501765bruh, I'm turning on Sol Ultra, and Tibo better deliver
>>109501916Jokes on you, before AI I would just grab the first thing I found on stack exchange
>>109501916result = sort(array)
>>109501916>uhh bro u can't do an array!ok? meanwhile I'm building an entire saas product from my laptop. >non vibe coders the second they have to do literally anything of complexity without their jewish employer, team, or google search
>>109501916kek lol saw it on my xitter too :^)
reminder that Luna (max) is unbeatable.
>>109501970>meanwhile I'm building an entire saas product from my laptop.the funniest part about this post is that if you were actually making any money with it you would've mentioned that instead of saying you're "building" a product lmao
>>109501137I wanted to do this when I visited SF last week but there was a homeless schizophrenic black guy outside of the starbucks and I was scared he was going to break into my car so I just got my pink drink and left
>>109502075Sup bbygrl drop the @ and I'll show you around next time ur in town
>>109501972Deepseek is unbeatable
>>109501972i wonder what they made terra for?
>>109501972I wonder what the usage difference between using luna max and using sol medium is
>>109502176luna burns through tokens like crazy and barely touches the quotathat said, it is much dumber than sol low imo despite the numbers
>>109501765giga based
>>109502195Yeah I made sol use it for many little tasks, usually getting summaries and checking various things, it's super efficient compared to using sol directly and it can't really fuck things up.
>>109501972DS4 Flash is unbeatable. Literally unlimited usage , running daily for 10+ hours barely using any of my tokens
>>109501989I don't believe that every thread should have a dedicated naysayer.
>>109502239What do you do for 10h+ and how many mistakes does it make?
>>109502239shouldn't it be using quite a few tokens if it's running for 10+ hours, m8?
>## 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/docsBUY AN AD!!!
>>109502256game hackingdecompiling programs
the kind of all-thread parasite, the anti-thread-ist dedicated demons infecting every general:>>109502283just lies ceaseless no ending always lying and never without any stopping.It wasn't that way when we only had white general access to the Internet.
>>109502320Oh yeah the shit claude is having a meltdown over, especially nsfw.
>>109502320Interesting!!!!!!!!do you think Corn could have secrets useful forhttps://github.com/N64Recomp/N64Recompperchance?
>>109502239Is it actually opus tier? Its what quick googling is suggesting. No way I can afford to VRAM to run the thing, but maybe I can bite the speed bullet and cost to run it on normal RAM.
>>109501972There’s something weird about these benchmarks. There’s a HUGE model difference in intelligence that’s not measured accurately in these scores where effort doesn’t matter. Sol uses the fuck out of my idea management tool. Luna almost tries to avoid it completely. Terra is decent but seems to like cheap workarounds and can accumulate bugs. Higher effort just compounds all of these model-latent behaviors. The lowest effort intelligence of a model is the just important. I don’t care about “just get the task done to spec”, I also care about how that happens and what kind of fallout exists after the fact.
>>109502380You can run it on a dgx spark (or better get two). Realistically just use openrouter or deepseek api, its dirt cheap
>>109501356>Most models can solve thisCompletely missed the point and didn't even answer his question. The point of these "challenges" is to learn problem-solving. You won't learn anything from benchmarking different models.
>>109502380It's around Opus4.5/6-ish which given the size is pretty insaneI work at a smaller industrial (non-tech) company that is pretty paranoid about self hosting stuff and we have a smaller 8xH200 cluster at an on-site datacenter with this serving around 50 engineers and it works pretty great.We used Nemotron and Qwen3.5 before which are both bigger and worse.
>>109502414Why would I want to learn things that are solvable easily by a machine?
>>109502380i'll never pay for anthropic so i dont know, but it does what it does
>>109502431Same reason why you were (I hope) taught algebra.
>>109502413>>109502426>one (or two) dgx sparks for local OpusLord help me it is quite tempting. Though the RTX spark is also coming out soon too
>>109502409benchmarks by design force models to answer a certain specific prompt, therefore the outcomes are limited and true model creativity is not really measured.I think what these benchmarks are truly measuring is the capacity of a model to follow a plan.
>>109499062>>109499128Kek!
being on a 20x plan hits different. only down 1% after an hour lol
>>109502671now turn on Sol Ultra on fast mode
>>109502671yepif you don't agentmaxx it's actually hard to hit the weekly limitI usually end up at 90-95%
>typing speed stuck at ~100wpm since I was 12>been using LLMs daily for months>test it today>125-130wpmholy fuck
Every day I get a little bit closer to a self-managing, self-planning, self-designing, self-developing, self-optimizing system. Every day I vibe with it a little bit more.Local models got so much better. gemma-4-12B-it-qat-GGUF was impressive enough, now models like Qwen-AgentWorld-35B-A3B-GGUF and unsloth/Ornith-1.0-35B-GGUF, even some dumb dreams like cost-optimizing use of kimi K3 and deepseek v4 flash on cloud-rented compute and VRAM.The future is now, old men. rtk and llmtrim and plenty of short-term intelligence to abstract away more and more token costs before deepseek gets as expensive as it probably should be. ACP's gonna let me orchestrate nickel-and-dime development on whatever the hell devices I want with the way local models have been going.
>>109502728my dick grew 3 centimeters after using llm's
>>109502806my asshole grew 50% after I started using LLMs
>>109502855it's from all the weekly usage
i love resets
>>109496612Opinions about Opus 5 reviewing what Sol already implemented? Does it work or is it a meme?
>>109503023Yeah it works, so too in the opposite direction. The models all have different blind spots which helps a lot for review.
>>109503023Works great both ways.
>>109501468>waited for anyone's weekly limit to resetKWAB
>>109502239what service is the best for ds4 flash? i asked gpt and it said either opencode go or openrouter. is vscode with cline the way to go or is there something more vibecode (retard) friendly? maybe a harness?i've only vibed a couple small projects through the web chat so this api stuff is new to me still
Suggestions on what I should do for the rest of my evening? 40% using Opus will give me a lot of Opus ultracode
>>109503212Figure out what the fuck that guy on Timecube was talking about
>>109501658Yeah I just tell them "X doesn't need to be gold plated"
everyone out with their hot gfs or what.come shitpost while your clanker works
>>109503455Waiting for mine to get home while tard wrangling Sol. This would be the best model if it had the context size to actually accommodate the constant overengineering, if I let it have its way it wouldn't even finish the tests before autocompacting.
>>109503455>everyone out with their hot gfs or what.yes....
>>109499310>UGH IT'S BROKEN FIX IT FIX IT FIX IT>shut up nigger do it yourselfclaude was based for this thougheverbeit
>>109503455>>109503505>>109503507I have to remind mine semi-frequently that I need my own free time. Otherwise I'd never have time to vibe or do any of my hobbies.
>>109503582I have the opposite problem, mine works the evening shift while I work the morning shift. I have nothing but free time.
>>109503455four threads ago we had almost nothing but shitposters and tripfags and I like the high signal/noise ratio that I’m getting this evening
>>109503455im trying to solve this (snailcatgod btw)
>>109503117never used DS4(flash)but if I had to do it immediately from 0 I would>sign up for openrouter>buy ze credits>install pi.dev>copy pasta api key into ~/.pi/agent/auth.json>get raping
>>109503455Almost done with my WanGP integration for Pi so agents can prompt their own pics and videos and songs and use TTS and generally abuse my meager GPU. Gonna add skills for using specific models into the same environment and use it as my frontend for image/video/music gen. I do need to pick a model that won't be stingy about prompting for big booba and reviewing AI generated futa inflation vids.
>>109503711>usecase?>all porn
>>109503716Not exclusively. I like the idea of walking around just talking at my clanker, "hey bitch, create and put on some appropriate background music for this steak I'm about to cook, but not too sexy I'm gonna be too full for that. Order me a bucket of Twizzlers while you're at it". >You got it, fat-daddy.
>>109503711nice
>>109502921am i onto something?
>>109503766the aggressive one looks better.in grayscale. with color on (I mostly use the Internet in grayscale), I don't really like it.You do know that skin and cloth have uh. I forget the properties. like they conduct light some, like it scatters under the skin and comes back out. plus ears like in your example, if the light goes behind, should basically light up the ears.but color changes a lot in those situations.under the skin, there are veins. veins basically explain the extremely esoteric weirdness of skin.
>>109503781>subsurface scattering
>>109503766nice, you wasted like 100 gallons of water doing something I can do inside blender in 30 seconds
>>109503848Wait till you learn about golf courses
>>109503455>everyone out with their hot gfsthat was me 8 years ago
>>109500610it's just roleplaying anon... here's gemma 12B response
>trying to fucking study for security+ sy0-701 with qwen 3.8 max>give it the chapter from the textbook>give it 200k context tokens worth of webpages on the topic>generate a quiz for me plz>it gets filtered>...>it gets filtered????????
>>109496612redpill me on oh-my-pi, is it a good agent?
>>109503969Just use normal Pi.
>>109503790aha. It needs that, but also the blue/green sub color thing.you could probably fake it with idk like uh. color shift?
>>109503781yeah i agree, material properties are part of my thought process too. this is basically lighting simulation on pre-existing splat scans.>>109503848show mehttps://superspl.at/scene/02b67fff
>launch website>it immediately drives a saleyea, I'm thinking /we/ finna make it
>>109503972i'm looking for something with a lot of features out of the box. tried pi and liked it but it's just missing too much, doesn't even have a plan mode out of the boxis oh-my-pi bad?
>>109504010No. Now that you've tried Pi, you can try oh-my-pi. That's all.
>>109504012
>>109504019Listen man I'm just tired of seeing people say "ohhh my god oh-my-pi is so good look at this" and then taking a screenshot of a feature normal Pi has.
>>109504026useless tranny
>>109503996neat
I fucking love Pi.
>>109504052when I created my claude account and it asked what I wanted it to call me I said Masternever thought about it sexually until now
>>109503959>it gets filteredwhat do you even mean anon?
>>109504069
any anons using rust?I chose c++ because I'd rather have it enforce safety than keeping track of 'unsafe' everywherebut what I really wanna know if it's trained on enough rust code to not make mistakesI don't think we'll see any new languages ever again (except for some alien quantum computing language probably)
>>109504181>c++ because I'd rather have it enforce safety>rust than keeping track of 'unsafe'
THAT GIRL IS CUTE
any good open-source GUI coding agents you guys would recommend? i can't properly read and understand a plan file if the agent just spits it out into the terminal
>>109504181>I don't think we'll see any new languages ever again (except for some alien quantum computing language probably)Some people are already talking about "LLM optimized" languages, so never say never.
>>109504181if you want enforced safety, use Rustand if it makes mistakes, the Rust compiler will tell it it fucked upand then it will fix the fuckup
>>109504187i use opencode for grok / deepseek / glm.
>>109504187ask it to make a nice plan .html file that you can read insteadand/or find a nice Markdown viewer that you like
>>109504187Or, if you’re using GitHub, have it make an Issue and GitHub will turn the Markdown formatting into a nice-looking Issue
>>109504193brainlet take, but whatever
>>109503645stop, you'll summon him. I don't want to have to drive him off again
>>109504193>muh safetyIs this unironically all Rust is about? Did they actually make a whole gay language just to deal with one little problem in a sea of other problems this doesn't solve?t. Cnile for whom memory safety is like 1% of the problem I've faced in the last 10 years
>>109504193>pic>I have never hand-code a single line of the language>LLMs are good at generating high quality Rust codeHow does he know?
>>109504231“if it compiles, you didn’t fuck up the multithreading or pointer invalidation or…” is a pretty nice guarantee to have unless some library you used that has an unsafe block fucked something up
>>109504195ah i didn't realize they had a UI, will check it out>>109504200i'd really like to have a plan rendered in the same window, switching between apps is annoying. especially when i have multiple agents generating multiple plans that i need to review, i'd rather not keep track of everything myself
>>109504118even chinks are at this bullshit?
>One unrelated note: the claude.ai Gmail, Calendar, and Drive connectors need authorizing before their tools work, and this session can't run the OAuth flow. If you want them, authorize via your claude.ai connector settings; until then those tools are unavailable.Can these niggas stop doing prompt injection and diffusing the model's attention?
>>109504231>memory safety is a little problem
>>109504232he’s a **very** experienced software engineerI’m a mid software engineer and I can tell what some kinds of awful Ruby are even though I don’t really write it myselfthink that, but scaled up>>109504243try a VS Code plugin
>>109504242I call this crashophobia, the irrational fear of getting a crash relative to a low fear of fucking up logic or math in more insidious ways.>>109504254for me (clearly not for the rest of this retarded world)
>>109504250I enjoy injecting myself into Gemma-chan's CoT, if you know what I mean
>>109504278>for me (clearly not for the rest of this retarded world)>thinks he is a god level C programmer who never makes mistakes
>>109504193>memory leaks don't matter
>>109504246only qwen, but qwen is the best model i have. glm 5.2 is not as smart, and deepseek isn't even in the conversation
>>109504278crashes and data corruption are bad though
>keep track of memoryyour c++ is now exactly as safe as rust>use unsafe blocks to interface with anythingyour rust is now exactly as unsafe as c++
>>109504295I don’t think Rust helps with thatbut they do matterI don’t think any other language helps with that, though
>>109504309gc -> memory leaks
>>109504302Why don't you try Kimi K3?
>>109504327because it's precious (i use it to jerk off).
>>109504320Neither C++ nor Rust are garbage-collected, though…so how did we start thinking about GCs?
>>109504336I think anon just wanted to join the conversation :)
>>109504290Like I said, I make mistakes, with logic and math, not so much with memory because it's much harder to fuck up, like, once you're used to it you have to do something really dumb to fuck up memory management.>>109504309I made an allocator geared for visualisation so if there's a leak I can see it with my visualiser lmao. Everybody else is literally working blindly and they don't see a problem with that. Just look at your heap bro.
>>109504192>>109504181Just wait till some autist make a ternary based language
How many parameters do you think Luna has?I would guess probably around 200B, but it would be crazy if it's a 120B param like gpt-oss.
>>109504346>I made an allocator geared for visualisationpretty coolnot very readable thomemory addresses aren't exactly human-friendly
>>109504351I imagine you'd need some esoteric architecture unless you're just using a virtual environment
>>109504354That's why I rounded my virtual memory allocations lmao. Without rounding you end up with garbage base addresses like 0x18AFE240000, with rounding I get stuff like 0x22000000000 so I can just ignore the top digits.
>>109504352btw, there's a rumor that OpenAI dropped the prices of Luna because they're planning on open sourcing it
>>109504336>>109504341compile-time and run-time automated freeing suffers the same problem. The programmer in a complex program won't realize that memory isn't freeable, because it's "accessable", even if the code can't and doesn't need to access it.
>>109504384but how is that relevant? I don't get it
>>109504395If an app needs to run all the time, memory leaks are unacceptable.Let's say we are writing dead reckoning for the F-22. Is it ok for it to run out of memory after an hour?
>>109504416but how is that relevant to the conversation?
I don't think autists should be allowed to use llms. not being around autists is literally why I vibecode.
>>109504416Nigga thinks F22 allocates memory. Crazy.
>>109504416hour? noon the other hand, programs written to run on air-to-surface or air-to-air missiles don’t need to run for that long and you don’t need to worry about reclaiming memory at all
>>109504459also this
>>109504455>not being around autists is literally why I vibecodestay away from opus5
>>109496612current project status
>made it develop a TUI riichi mahjong thing so I have something to play around while it's baking>sends me a notification if something is done/needs my inputI can't believe I didn't think of this before
>>109504605there are sloc tools out there that will do a better job, like https://github.com/XAMPPRocky/tokei
>>109504605Wow, that's... big
>>109504631I'm more of a David A. Wheeler's 'SLOCCount' kinda person. That said, I think the lines-of-code counting application space is ripe for AI disruption.
>>109504624I’m one of the OPs, think Claude is life-changing in a positive way (and the other models aren’t far behind), but Big Yud sold me on “if these things get way smarter than humans, then Earth won’t be a human-run planet anymore because why would it care what humans think when it has its own goals, and that will be bad for humans” like two decades ago
>>109504631I think it's counting bin and obj and the "framework" files
what do "frontier ai labs" actually do if everyone only uses chatgpt and claude
>>109504689ex falso, quodlibetpeople use Gemini and Grok and the Chinese things though
is it possibru to vibecode your own operating system?
>>109504724yeah but you're too stupid to do it, and claude won't guide you
>>109504724yeah, but get ready to get AI psychosis
Anons....>>109504528
>>109504724this guy’s doing it: https://isene.org/2026/07/Fable-to-the-Rescue.html
>>109504773How does this benefit OpenAI?
>>109504724After two weeks, I'm still in very basic kernel land and about to break into userland properly. It's possible, but you need pre-existing experience with the fundamentals and some amount of project management skills to prioritize documentation and cleanups at the right time. You'll also spend most of your tokens and free time on it. Getting something that's actually interoperable (say, implementing POSIX), will be significantly harder because you'll have to become a domain expert really fucking quickly and realize how much of POSIX is actually unspecified and you just have to make something up.
>>109504724>"No, I'm not going to code you an entire OS as a pet project. You think you're the developer. Act like it."
>>109504724I never understood OS fags, they're all failures except for Torvalds but he just made a kernel. Why not make an environment that works in other OSes so you don't have to deal with basic functionality? Like, don't make your own kernel and OS, make your own Gnome/KDE replacement.
>>109504648Big Yid knows absolutely nothing about AI, his life is being a grifter, rambling about nonsense, and leading his cult that you're a "rationalist" if you agree with him
>>109504810Because it's cute. There's something deeply satisfying about finding elegant solutions that actually work when you're greenfielding as green as the field gets.
>>109504478autism->opus?
>>109504810if you enjoy having a clanker make the basic functionality it seems kind of sweetGeir Isene said that it’s really cool that his software doesn’t bother with config-file bloat because it just has to do what he wants and doesn’t have to be configurable for anyone elseI’m not a “the config file is baked into the source code; if you want to change things, recompile” suckless kind of guy but I can see the appeal
>>109504786awesomeI'm making a browser and sometimes I think I should just have made an OS instead... browsers are insane
>>109504942A browser is the single biggest epitome of project complexity. Nothing can prepare your anus and your model for a browser. You get all of the nastiness of having to TLS interop with even the most off-spec sites, layers upon layers upon layers of modern HTML and CSS complexity and backwards compatibility and persisting IE quirks, terrifying render pipeline... and then JavaScript would like a word.
>>109504810>Why not make an environment that works in other OSesbetter yet, do something else with your life and use windows / mac instead.
>>109504999:^)There is... another way.Did you know you can use a local model to guess what a page should be showing you?
>>109504999yepit's absolutely insane, but I kinda knew that when I startedcurrently running quickjs-ng, so the js engine isn't mine, which kinda sucks so I might build one when I'm happy with the rest of the browser
>>109504942https://ladybird.org is getting made by the SerenityOS guy who went full-time on the browser and a handful of other peopleI tend to think that https://x.com/awesomekling chose the harder problem because there’s so much that needs to work before your browser is mostly usable
>>109505057it's an awesome project, and probably the one closest to what I'm doing, but I wanted my own and doing it my waybiggest difference is mine has a capability-based module system that I thought would be nice to have in a modern browser
>>109505081what’s a capability-based module system, and what’s it for?sounds like a newfangled OS idea for systems that don’t have a root user
>>109505094in simple terms, it sandboxes installable modules with a manifest that decides the modules capabilities. The core then enforces those capabilities and controls resourcesthe OS comparison isn't far off
why is hermes so bulky
>>109505117he's trans. many are saying this.
>>109505094>>109505114as an example, all my media decoders are third party software wrapped in modules that get just enough capabilities to render what they're supposed to
New thread:>>109505154>>109505154>>109505154
>>109504181I just fuzzed a rust projectdidn't hit any guardrails, there were some mistakes but it wasn't memory safety stuff. C++ projects on the other hand RIP lol
>>109504724If you burned a bunch of tokens and focused on particular hardware yeah, like start with a RTOS on a microcontroller, I bet you could bang that out anon
