A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs. -- Frontier models - start here if you have $20 or so https://claude.com/product/claude-code https://developers.openai.com/codex/cli -- B-tier https://x.ai/cli https://platform.deepseek.com -- Prompting / context / skills https://arps18.github.io/posts/claude-code-mastery/ https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/ https://github.com/mattpocock/skills -- Other editors / terminal agents / coding agents https://pi.dev/ https://opencode.ai/ https://cursor.com/docs https://osaurus.ai/ https://docs.cline.bot/ https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent -- Benchmarks / rankings https://www.tbench.ai/leaderboard/terminal-bench/2.1 https://artificialanalysis.ai/ -- What we’ve done https://vcg.gitgud.site -- Previous thread >>109579552
https://mistral.ai/pricing/For Europeans by Europeans.
>Grok is experiencing issues. We are working on restoring service as quickly as possible.
>>109582167call me when they release something good again
first for omp needs to be in the fucking OP
>>109582167The uncensored Qwen3.8 27B I put on the vast.ai instance is churning some good MITM hacking scripts for my Adalm-Pluto I got connected on my server :)
>matt pocock>simon williamsonwho the fuck makes this OP? which retards are responsible for spreading misinformation? This is just bad advice, don't tell me that the website you guys made is also full of this gobbledygook
>>109582190Only settler colonialists use grok
if I want to build a game inside the browser based on another game, should I employ sol or fable? which is more capable in the sense of game building? I've heard that codex models are better in the art generation department so maybe I should use sol
>>109582328Neither are good by default, what make them good in game development is if the user can actually program or not. If you can't you will make a steaming pile of garbage and you will be wasting your time.
>>109582349oh dear, you sound like a luddie. >>>/dpt/ is down that way leather boy
>>109582197hello glowie
>>109582373I've been experimenting with LLM generated code since the day chatgpt dropped, I'm so much more experienced at this than you could possibly understand.
>>109582393>been experimenting with LLM generated codeSounds like a snailcat to me. Do you vibecode, or nah?
>>109582197Cool, but I thought cellular networks required way tighter latencies and much higher bandwidths than that?
>>109582378You have to go back.
mega based.
>>109582393sure you have. post an AI assisted game you've made since you have more experience than I could possibly understand. let's see it
dumb thing I did but I wanted to exploit the web ui since my codex is always hitting its limitthe linux terminal container is honestly not as restrictive as I expected I had it copy pi agent harness and operate in the same manner and sent it into loops to try dumb things like install claude code and a local model that it was pretty easily able to dounfortunately it isn't very powerful no GPU and only 6GB ram dont exactly know what to do with this info but perhaps someone might have further ideas how to exploit web ui to vibe code and not have to shell out more
>>109582419I make money "vibecoding", I know how this shit works. I've made my own agentic harness. Cause I know how to fucking program. You guys have no sauce, your tools are far behind anyone who actually knows their shit has made. >>109582442I'm not gonna reveal my shit on 4chan you retard. You're beneath me. But I can tell you about, wgsl shader pipelines, webgl2 backpropagation shaders, three.js, raylib, wasm, whatever the fuck you want. If you don't know how to program you will make abhorrent garbage. You have to UNDERSTAND what you are building.
>>109582442Not him but why the fuck would you think everybody makes games? You know the game industry is a small part of the overall software industry, right?
>>109582453Well...you could break out the container with a loonix 0day and take over the entire datacenter your session is residing in?
>>109582479How do you make money vibecoding if you don't mind me asking? Do you just use AI in a normal software development job or are you self employed?
Wait a second, if every person in India is a super talented programmer to the point where India should basically control the entire IT infrastructure of europe and america, why hasn't India made a single AI model yet?Shouldn't it be India making all those open weight models instead of china?
>>109582499Probably because making good AIs requires access to millions of GPUs.
>>109582515I guess that's a case study in unlimited indians vs unlimited machines.
>>109582479actually lol'd reading that because I know you're being 100% serious hahaha. you wont reveal shit because you only made shit lil bro LOL. fuck outta here
>>109582479>gaymeGloat elsewhere about your furry hentai game
>>109582515The deepseek team has no resources at all
>>1095824536GB of RAM. are these containers persistent or simply persisting as long your loop runs? Crooked I is wondering how many buttcoin farming and bruteforce cracking tools can run if you create a thousand gpt accounts and do it on all of them. or any other botnet compute inference with low ram requirements
>>109582486>ask specifically about making a game with an agent>retard snailcat chimes in his worthless opinionnice reading comprehension retard
>>109582479>You have to UNDERSTAND what you are building.Luddite seethe is sweet. You are the exact same as artists reeing about generated art not too long ago.I hope you didn’t spend too much on your education and training, hopefully you didn’t take on a lot of debt and interest.>a self made person wouldn’t be nearly as asshurt as you are
>The orchestrator is asking me to report, but I haven't done the work yetkek
>>109582421Nah its gucci, I asked LLM >>109582378hello paranoid schizo that isn't self aware enough to understand his paranoia stems from struggle of the unconscious shame and shadow work you didn't do on yourself
>>109582529Yeah I'm sure a HFT firm is very poor
>>109582534oh my bad
>>109582556huh, interesting
>>109582547he's just like me
ben did the thinghttps://youtu.be/U2Mw9MS84DY
>>109582660Please stop decreasing the signal to noise ratio of this thread.
>jannies got replaced by AIWorking for free was not enough.
>>109582498You can be employed and vibe code, most programming jobs are piss easy with the most in distribution stacks. If you're good at it they let you take over whole projects by yourself. Because, A, you don't actually vibe code the important stuff and B, aren't retarded.>>109582522>>109582525>>109582542classic cope seethe posts, you have nothing, made me smile
Copex turds getting rugged afain
>>109582612the report is nice but I wanna get into A5/1 crackin soon wiggaaaa
>>109582747
> 4 months later> Mythos/Fable still king by farwhenever OpenAI drops their Astra thing, Anthropic will answer with something even more amazing
>sir, you have too many transactions (paypal)>your account is suspicious so we're holding your moneylol??? what the fuck is their problem. I sell digital items, why the fuck does it matter that there's no tracking number.
ran out of usage everywhere so im vibecoding with chatgpt
>>109582869>so we're holding your moneyKEK, good luck getting your money back from ((them))
>>109582917It hurts, because a really generous user just sent me $100 (can hardly fucking believe it myself, the app is only $5).
>>109582869I think that is pretty standard when you first start making money on paypal? Hopefully they dont drag it out on you too long
>>109582869i really gotta move my money from paypal before they do thisits like 20 years old and in my moms name so they will most assuredly give me shit if this ever happened (she has been dead 10 years now)
>>109582869classic paypal move
>>109582869>he doesnt know about paypal shenaniganskek. they lock your money or fine you for wrong think. not even joking, look it up
>Make a calculated token gamble straddling the weekly usage limits at 97%, should barely scrape by through the skin of my teeth>all going well>time comes to write the class>opus 5 tries to use the fucking bash tool to write it with the command line>its too fucking long of course, the bash tool fails>1200 lines of code worth of fucking tokens GONE>just barely enough left to write the class again and do half the tests, now all the memory is stale and it didn't do the "finishing touches" it said were remaining before hitting the weekly limit>now i have token debt to unfuck this piece of shit and cap it off when usage resetsI guess it's time to add "Don't use command line to write code" to CLAUDE.md, not that it doesn't ignore what's in there already half the time anyway.This shit is still a fucking token lottery through and through
Alright, I made it for fun because someone threw that OpenRouter Sol discount into the mix: https://cavedex.bradthomasbrown.com/You can type in your situation and the idea is that if you want to use Codex right now, it’ll tell you how to do that, or if you expect to run out before the usual reset, Tibo shenanigans not included.For me, it’s actually most worthwhile at my pace this week to just upgrade my subscription or buy another $100 subscription rather than use API credits, even if OpenRouter has them at “up to 75%” off.>TL:DRAPI credits are still a retard taxI can update the page if there’s interest so it provides a schema, your clank can fit data to it and it’ll be more tailored towards your exact usage; a lot like the cube can take data.
Gotta say, you get quite a lot of grok on Cursor $20
You know you're having a good night when you're pushing Claude to doublecheck something he did and then he starts talking about "my vandalism."
>>109582742>retard's only response is "seethe"lol
>>109582747Whatever they're doing is thousands better than Misanthropic. I cannot use Opus or Fable anymore, there's literally no task that I can think of that would be worthwhile and still fit inside the 5 hours limit. Zero.
>>109583585Wish I could see the real CoT for this line.
>>109583701The thing is, for Anthropic, your worthwhile tasks should be making money so you can pay for the actual sub and not the x1 or x5.
>>109583708Used to pay for the 20x, dropped to 5x because the experience got worse.
>>109583708/s? Making money with these tools should not be the only reason to use them.
>>109583753Oh, absolutely, I should clarify. I have a few tasks that make me enough money to pay for the 20x and then I use the rest of my quota on stupid bullshit.
>>109583701you're retarded if you let agents run wild for over 2 hours at a time and dont build things modularlycant imagine the piles of overengineered trash that you prompt
>>109583761>2 hoursIf it's a task that needs context, you can run through the 5 hour limit in 15 minutes now. Easy. Even without Ultra. It's not a matter of letting agents run unattended?
>>109583778this means you're bad at managing context + your shit is so overengineered you need way too much context as wellsounds like your shit doesnt have a clear separation of concerns, your memories are bloated, and you don't know which to tell it to pick and choose from, compounding the issue
>>109583798Please stop trying to instigate a flame war, not the time.
>>109583808t. codebase wide spaghetti task bundler extraordinaire
>>109583585>>109583702Damn, I just told it to spawn a Opus 5 subagent and it actually got figured out. I always assumed it was total benchmaxxed slop not worth using but it got the job done.
>>109583808It’s an angry anthropic fanboy that thinks he’s a genius because his shit company makes shit models and has shit pricing, usages, and quotas, and he’s “figured it all out” in how to deal with being treated by human garbage by pushing out tiny little bits of code in his 5 hour cuck windows.
>>109583872NTA, I genuinely do not understand how you guys hit limits. I've never hit a limit on Codex, I hit the limit on Claude exactly once.
so it seems like both codex and claude code are rugpulling and reducing usage limits and making models retardedi just fucking cancelled my claude 20$ and bought GPT Plus and the weekly limit magically goes down 5x faster than it normally shouldhow fucking hard is it to get an Opus 4.8/Sol-tier subscription that won't nerf itself in the middle of the night?
>>109583890I hit the limit on codex AND kimi on the $200 plans in 3 days using Sol medium. I do use the 372k context setting on codex though which consumes more usage.
>>109583691thanks for the laugh buddy, you'll get em next time ;)
>>109583953>luddie still lurking the thread for gotchaslol. wonder if you realize that someone who actually had nothing to prove and was mentally stable wouldn't do what you're doing
>>109583940The response is to use a local model (large upfront cost, potentially doesn't keep up with latest releases), or API from a major cloud provider that let's you use a specific version of a specific model without having to fear A/B testing or unexpected attempts at increasing efficiency by making parts of the experience worse (no upfront costs, but large recurring costs and although much safer from shenanigans than with subscriptions, never *entirely* safe from them).
>end up creating a keylogger without realizing it>claude hasn't flagged it because I was building a type racing application.
>>109583980What are your arguments? You can keep saying luddie but it doesn't mean anything, say something, anything. I'm here to see what the retards of vcg are doing, you're all behind it's a pity. Where's the alpha?
Not even halfway through the week and I'm at 73% usage If this limit is "temporarily boosted", how fucking bad is it usually, jfcFuck you tibo you're definitely not getting another 20 bucks you cocksucker
>>109584004>was thinking of switching to codex because of this exact problem with claudeMight just hop to Open code at this point.
>>109583989>say something, anythingpost a project of yours
STOP USING CODEX STOP USING OPEN CHODE THEY ARE GARBAGE HARNESSES THEY WASTE YOUR TOKENS
>>109584014That's not an argument. Why? I could but it would take a bit to deanonymize, maybe tomorrow, I could make you something fresh. What do you want to see? To learn? What questions do you have? Why don't you post a project of yours?
>>109584015What I always hear is that Claude Code is the inefficient one. Granted, it's hard to compare since there's no earthly reason to use other models in Claude Code, and Anthropic charges you the extra usage rate if you use Claude with third party harnesses (unless you do it in a way that violates ToS, but then they can ban you, no refunds, so...)
>>109584015what is the alternative then?
>>109582138where teto snailcat
>"you're beneath me. I'M the alpha here. I know how to fucking program. I'm MUCH more experienced than you can possibly understand." >okay, post a project so we can see>"n-no..."
>>109584047Pi? Potentially the new DeepSeek harness?
>>109584042>>109584047CC is poopoo, I've taken a look at the source when it leaked. It is so mega ass and I can't even imagine how much worse it has gotten. Use OH MY PI (it can spoof CC and its not openclaw retardation so it falls within TOS atm). Codex is way behind omp, they are basically just copying features that are months old atp. If you don't want omp, use pi and customize it using itself. Both are customizable heavily in very simple ways. Omp has an extremely good edit tool that uses less tokens than other harnesses, and that's basically what you are paying for the most, so why pay extra? The deepseek harness is interesting, need to look into it more, rather a purist idea. There's also prime intellect agent with the ipython kernel mode (which omp also has) and the autolith lisp based one, both interesting projects taking a look at something other than bash. I've been studying and using virtually every harness that has come out in the last six months.
>>109584060You're only response has been to pretend to be laughing on the internet. (you) have nothing to say sadboi ;(
using free agent feels so good. free token, free hardware. free anything. still can accomplished thing.
>can't use claude in a 3rd-party harness>but can use literally anything in CC through a proxyyep, CC wins again
>>109584113>...
yeah im convinced you guys are all shill bots or actual nocoders who have no idea how to direct extrapolation or debug and just say shit like do the thing, make no mistakes
>>109584113>can't use claude in a third-party harness
>not being able to program is good when using the thing that programs actually
One thing that's frustrating with Claude is how the paraphrased reasoning trace sometimes makes mistake. So even if you pay attention to it to follow the model's reasoning, it will sometimes make it seem like the model is getting badly sidetracked, so you stop it to explain, only to realize that it's just the summary that misinterpreted things. I'm noticing it with Opus 5. Surely that's the case of it being a cheap model that's used for this task, and not Anthropic being hostile to users to the point of poisoning even the reasoning summaries, right?
>>109584151>Plan/Scaffold with very high effort setting>Drop it down to medium to flesh it out in a fresh context>medium finds and fixes a bunch of mistakes very high made when it tripped over itself because it got lost in its own sauceMany such cases. It's a coin flip for me at this point whether I use anything higher than medium
Ok guys, here's a lesson for y'all. Do NOT use quantized inference when doing reinforcement learning. I tried switching my inference server to Q8 and guess what, it made the model collapse and costed me 50 dollars in compute.
>>109584151>>109584174just use 4.8
>>109584174My weekly usage resets in a few hours, so balling on ULTRA right now. I asked Claude to review/plan some small thing, but the reasoning summary made it seem like it was taking into consideration something that seemed, at least to me, completely irrelevant. So I stopped it to try to redirect, and it's response was "what you quoted isn't from me", and that although it had briefly thought about that thing, it had been in a different context.
>>109584244do a memory audit
>>109584257Manually?
>>109584272both, incl stipulations that you want to prevent bleeding irrelevant context
are the snailoyim gonna be alright?
Hi, I'm stupid and trying to get into this. I installed Ollama and through it qwen3.8:27b then got OpenCode pointed at it. It seems pretty alright, but I'm wondering if there's something better for what I've got (a 5090), or should I just keep on with this?
>>109582486oneshotting simple games is one of the oldest and earliest llm benchmarks
>>109584197You were finetuning the bf16 version while using a quantized version for rollouts? At what frequency were you reflecting changes in the quantized version? It doesn't seem ideal, but intuitively SFT would be even more out of distribution and still work, no?Since the pic is called bf16-recovery and the first picture shows it recovering, or on its way to recovery after that huge spike, is it possible that something else happened?
so do you study the code the AI produces so you know how it did it or do you just feed it another task cause who gives a shit
>>109584419It depends.
>>109584358no that's the best model you've got right now
>>109584408It's not finetuning, I'm training a tiny transformer model from scratch through self-play to play a game similar to AlphaGo and such. There is no pre-existing data to do supervised training on.I've tried many things so the data is very messy.I switched to Q8 inference at around x=86 and only switched back to bf16 in that chart during the last 3 data points which is when the training loss in the policy begins to decrease. The big ending money spike at x=110 is me switching to a different search algorithm (the RL loop does self play with search using the value model which then trains the policy model).I'm doing 1500 games per training cycle (each game has 700 turns) and using each sample for 4 training cycles before discarding it (so 75% rollouts from older checkpoints and 25% with the latest checkpoint) so the total sample pool is 6000 games.One complication is that the ending money is related to competence but not really the target because the real objective is money relative to the opponent, so I don't have any automatic objective competence evaluation (I probably should even if it spends a bit extra time per cycle). Each cycle is about an hour and a half on a 4x5090 I think.
>>109582138is there a good bench that takes into account the fact you can chat with grok while it's thinking? Like it's designed for agentic, and so you can just basically be another agent.Like it's like "we should bla bla bla" and I go "yes, do it that way" or "no, I want to fully implement this".It's really a great way to save time.
So I'm not even sure the Q8 actually caused a collapse or it was just learning a different play style that wins less money overall but wins more, based on evaluations I did (not visible on the chart) it seems to not have degraded significantly on win rate during the ending money dip but the increasing policy loss seems to indicate something wasn't right.
ALERT ALERTpeak hours coming up in 60 secondscease all proooompting immediately!
>>109584743calm down, there's enough tokens for everybody
>>109584761Yes but you have to pay twice the price for the next 4 hours.So the smart thing is to use those 4 hours for planning your detailed prompts and then firing them off in off peak hours.
>>109584770I wouldn't know, I'm a subscriptionchad
>>109583798how do I get gud as a moron
It seems prime-agent is hit by one of pi's most retarded ideas: rewriting your chat history as a single string.Are people really stupid enough to use this piece of shit?>[User]: What they said>[Assistant thinking]: Internal reasoning>[Assistant]: Response text>[Assistant tool calls]: ipython(code="open('foo.ts').read()"); >edit(path="bar.ts", ...)>[Tool result]: Output from toolThe dev unironically thinks that this is a good idea. And I have seen him brag about being efficient with the cache. What a fucking piece of shit, just like his users. I hope they all die.
>>109584782more just a subcuckyou are paying for tokens you aren't even usingor are you telling me you max out every token limit 100% every single 5 hour window in a month?
>gearing up to go public soon on my app>setting up the domain>there's a .org company with the same name in the same field that looks like it's from the 2000s and hasn't been used or updated in yearsit's so similar that if someone googles my app, they are going to see two results: that and my thing. and they will absolutely get confused which is which. I'm really bummed because I loved the name
>>109584843>you are paying for tokens you aren't even usingnta but I'm not.
>Same mechanism verified sound under the window manager, not just the compositor; no new bug found beyond the already-verified-suffix bug. oh my fucking gooodddddddddddddd
>>109584843Lol I could probably use %10 of my subs and get a better deal than API. But yes, I use the weekly usage from my Kimi and codex $200 subs in less than a week. I don't get close to using all of the Opencode Go usage but I use more than I could get for $10 through API.But honestly next month I'm probably going to scale down to $20 OpenAI just to use ChatGPT with the MCP trick, $10 Opencode and I haven't decided yet whether to keep the Kimi one.
>>109582115I just talk to it lmao. Tell it what I want to do, tell it to show me its plan and ask questions, answer them, then tell it to proceed, then test, ask it to change things and so forth.
>>109582913how does it work?
>>109585131https://desktopcommander.app/mcp/chatgpt/
you can use grok to grok documents. should've been obvious. I have it create an md of processes, but like there may be a better way.
>>109585152Based, thanks! It's like having double the usage now.
"just learn a trade, bro" retards should watch this videohttps://www.youtube.com/watch?v=agi3xLTGyaU
holy shit I'm 3 days into learning rust and I'm about to give up I don't give a fuck anymore if it's like 10% faster than java, doing literally fucking anything in it is so fucking painful holy shit. so much padding and bullshit to write because the language won't implicitly convert anything, so much retarded bullshit around "borrowing" references or whatever the fuck, literally every "class" method requires writing self.whatever the fuck 100 times, even fucking initializing a struct is a fucking pain. fuck this.I just want to code and get shit done I don't want to be fighting with the compiler because nothing makes sense.
Why haven't you retards made an exhentai uploader yet
>>109585306That's why only deranged trannies love it
>>109585306Are you writing it by hand? This is /vcg/ you only need to know enough Rust to review the broad strokes if you are using AI.
still 1 hours and 15 minutes before we are back in off peak hoursi can't take this anymore i am literally shaking.. i NEED MY FIXi tried coding by hand, and sure i can make progress, but i feel like a fucking snail, i do in 10 minutes what takes llm 10 seconds I am completely brain fried. I cannot stand to exist without tokens
>>109585355Not him but if you want a job you have to pretend to know how to code the snailcat way.t. trying to fake it to get a Qt/QML/modern C++ job
>>109585152The future is renaissance dudes who can do everything.Guys who can do blue collar, white collar, pink collar, green collar and wear a suit.
>>109585306Using rust is easy. All you have to say is "hey use rust to make the website" and the AI makes it.
>>109585397>i tried coding by hand, and sure i can make progress, but i feel like a fucking snail, i do in 10 minutes what takes llm 10 secondsskill issue
>>109585397Yesterday my friend struggle with Qt in C++ and I was able to explain everything perfectly. Feels good.
>>109585495>>109582138Disgusting. Why are you wasting tokens making waifus waddle in shit?2D deserves better than to be thrown into a pajeet shitspace.
>>109585561There is no shit in the first image and she's cute and hot.
>>109585569You could smell the shit from the streets below, pajeet apologist.Thousands of places that are beautiful.Hundreds that use sari or an equivalent.You chose the worst one.
>>109585491ok but how about leather collar
>>109585577meds
>>109585589Fuck off shit eater.
we're so back!
>>109584419I don't bother at this point.I just have a separate hand-coding session and an AI session where I only give a quick glance at the codes.
Any poorfags here?I assume the best combo you can get, if you are destitute and desperate. Is probally Mimo 2.5 monthly plan, make your own harness with the langchain dcode and strip stuff out - only use the 'interpeter' i.e. where the AI can just code instead of native tools. People say Opencode GO is good, but they trick you. They charge 4x what you get on deepseek API, then people all say it's a good deal because they never check against the actual price. I think this is my plan. I am out of money, and I have to finish this project. I spent 100 dollars on deepseek flash and claude code, and I still have to do so much
>>109585510yeah i mean i understand al the code the clanker writes, i could write i all by myself too but it simply makes no sense.. it would be like using my hand to dig a hole when shovels exist
>>109585659are you sure? i run out of kimi and codex on the $200 plans, but I can use ds4 flash all day on the opencode go plan. you pretty much cant run out without using multiple agents 24/7.are you sure you are not paying for opencode ZEN?
>>109585674not owning the code is the best feeling i have had for years
>>109585674I currently also don't write any code, just feels good to know my brain hasn't fully atrophied.
>>109585680Never mind you are right. Same pricing.The best deal must be to get the opencode go plan, then get some stripped down harness.I don't know how people use oepncode though, it can't background terminals at all, you just sit there and wait. It's so slow.
>>109585708Yes, I bet this is the best deal you can get.You get opencode go plan.You get dcode, get an AI agent to vibecode it to only use the Interpeter tool mode and not raw tools. I think dcode can background tasks. Sucks being poor, but we must find a way
I'm trying the meme
>>109585708you can use any harness you can with it, they give you an api key, you aren't forced to use opencode.
>>109585719>get an AI agent to vibecode it to only use the Interpeter tool mode and not raw toolswhy would you kneecap your already retarded robots instead of just using a normal harness
honestly ds4 flash is good enough if you aren't on a time crunch, it'll just take longer. but probably not even twice as long.
Is there even a good harness where you bring your own api key.I tried opencode, can't background terminal. I tried claude code, but it seeems to use my money too fast.I don't have to vibe-code it I don't think, maybe just a small amount of optimization. Langchain is a big company, they made dcode - it comes with a TUI. Desperate times call for desperate solutions. I wrote a whole sever in python I need to move to javascript, it is a lot of code.
>>109585727stop posting about that faggot and his faggot useless project here. I'd rather forget he exists.
>>109585751meant for >>109585730If you have a good harness, please share with us all. I am just experimenting here
>>109585751you know modern chatbots have python sandboxes, right? you could upload your codebase as a .zip and get the code ported by chatgpt or any of the other ones. heck you probably can get kimi on the web for free and it probably can work with zips.
>>109585757I didn't post about him I'm just trying his meme app.
>>109585772>>109585729I will try these, thank you
Are there any good courses or articles I can use to get educated on customising and steering Claude for use at work? Auto mode scares me and the times I turned it on it went to do something I didn’t want it to do, but reviewing commands one at a time slows down work so much
I've been out of the loop for a year or so. Used to proudly be anti-gAI but it has caught up with me and is likely the reason why I never managed to land a job out of my education...After NEETing for a while and just hating life man, I've made the decision to give it another shot and hopefully land a job this time around. Haven't written a single line of code in months, would you recommend me to opt for claude pro or some other vendor? Will 20$ rly get me that far or is it a scam
>>109585853Depend on what you build.I use the 20$ for web shit, and I barely run out of token. I imagine if it is much harder and token-heavy in game dev
>>109585751just use pi
>>109585853Claude will suck you dry and their cheap models are garbage, only worth it for their big boy models on the most expensive tier. Buy ChatGPT $20 Plus sub, download the desktop app (formerly Codex, now ChatGPT), try to use gpt-5.6-luna (the cheap model) for most things, switch to gpt-5.6-sol for problems that are harder
>>109585561Web chat image generation is free
>>109585923I tried that one, but to get core features like subagents you must get some random person on github. There's also multiple version of core packages. They built off each other too. You can't tell what is jank, and what is useful, or when the dev will just roll up and soft-quit.
>>109585942oh you're lazy and entitleduse opencode or codex
>>109585959Lazy? Okay. Yeah, I want some pajeet from India asking opus to make an extension that OTHER people build on, then watch it all collapse as I need "big boob ant" extension as a dependency on shit I need. How do people have faith in random people making their harnesss
>>109585973so lazy you can't even conceive building the extension yourselfmust be european
>>109585973Let's spend hundreds of dollars worth of tokens to make things that already exist and the hundreds of dollars worth of tokens to configure them
poor too.definitely european.
>>109585979>>109585982What is with the shitty Pi meme? It is almost as bad as OpenClaw. Bro, there IS no harness. Some guy just made a basic typescript, then a thousands jeets descending to make it semi-usable in the most jank way possible.Rashesh for terminal commandsPajoot for any visualsSareen for the delegation system.Wow guys! Great idea. It's too late though, I already won. Once I rig this thing to use code for tools instead of tool calls, I will be able to actually make money off this AI bs.
>>109586009yeah man, go for it
is this O(n)?
Hope this is the right place to ask but is anyone "vibe cracking" apps? I have a few old Windows 7 apps for small businesses that I want to dump and migrate their database or the odd $400 CAD plugin that I do not want to pay for. Best way to approach this? I'd ask Codex but its policy is giving push back (I didn't even wanna bother with Claude thats a certain ban).
>>109586009Hermes gets work done without nonsense.First I dismissed it because of the anime girl on the cover but it actually works.
>>109586059there used to be one anon in here cracking mobile shit to farm currencyi've only had luck asking for for decomp/recompthey'll happily rebuild the entire app for you and then make changes, but as soon as you ask about copyprotection they'll clam upnot sure if the chink models are like this - might be worth throwing a bit of money on openrouter and seeing if glm/kimi/deepseek hit guardrails
>>109586075Oh, that is a good coder too? That is great, I will have to try that and see if it can shoot out a lot of subagents
>>109586094It has a function to delegate to subagents, but I think the number of subagents is lower.
>>109586075i stopped using it after they dropped the minecraft and pokemon skills from the default configit's soulless now
So what is currently the best bang for the buck? WHere i get the lowest price per task?I don't want to spend more than about $15/monthI want a capability that is not worse than deepseek flash v4 at high effort.I don't care if it's a cost per token or a sub, i just want to squeeze maximum juice out of that $15 monthly budgetthanks frensGod i am so fucking pissed they raised the the DS api prices.. it was PERFECT for me..
>>109586150gemini 3.7 flash or luna max
>>109586150probably the $20 openai subsubsidised luna + sol when you need itprobably opencode for $10.chink subs like glm and kimi i think are about 18/19 month to month - i've never tried them for that reason. i'd dip in for a tenner.>>109586169i can't with gemini man - 3.1 - 3.6 were all lying, gaslighting, piece of shit models and i fired 3.7 up yesterday and just had no patience for it doing retarded shit
>>109586169is ds flash4 in that grapg pre or post price hike?since by those numbers it seems like gemini is a bit more capable but also significantly more expensive
>>109586205>lying, gaslighting, piece of shit modelsfunny.. thats what ALL llms act like>>109586208>updated aug 13https://deepswe.datacurve.ai/
Hey fellow europeans have you tried Mistral pro?It's only 6 euros for students.So I was wondering if it is worth it because they're charging money, so they must be competitively priced.
>>109586222nah, most models when told they've made a mistake don't insist that it's actually a problem with your test environment and waste your time by telling you go check shit. it wouldn't be so bad, if it was fucking wrong every single time.
>>109586205>probably opencode for $10.>probably the $20 openai subvery roughly in dollar value, how much would you have to pay in DS api off peak prices to get that same value?By that i mean if i sue up every token in the $10 opencode sub to do work, how much would that same work cost me if i did it through DS flash api?
>>109586234>if it wasn't*>>109586237this is what they're claiming right now + trying to match pre-price hike rates - are they quanting the model? who knows5 for the first month so worth throwing money at to test it yourself
>>109586205kimi is not a good bang for your buck, it is a bad one as a matter of factmoonshot makes anthropic feel generousplease stop misleading anons if you don't know what you're talking about
>>109586250>i've never tried thembut fair enough. i thought i'd seen a couple people with subs here say rates were decent.
>>109586248please god don't let them rugull is with quantized flash trash
>>109586237If opencode subs are anywhere as subsidized as OpenAI subs, API costs are 5x subscription costs (if you use the full subscription)I have no idea if they subsidize their subscriptions at all, though (I haven’t tested it myself), but a quick search suggests it’s maybe the sameSo>use a full $10 sub>it’s maybe $50 in API costs
>>1095862875x is hugethat would be basically same pricing as old DS api priceswill def look into itoh and i hope they won't force me to use some gay ass shitty custom ide. I am a vscode only giga chad and i intend to remain one.
>>109586270the only decent thing about moonshot subs is that usage increases non linearly with cost but the $20 plan is trash and that's the one you were recommending
>>109585852Um any tips for a newfag vibe coder????
>>109586248Seems too good to be true... and it is
what does scammy-man feed his AI?
>>109586293Try to test it, I’m just guessing based on my codex subscription. The $200 plan is actually disgusting, I’m going to be upgrading to it today. It’s 4x the usage as the $100 plan, or roughly 10x the value if you make good use of it over the API costs (OpenAI math)
Is it just me or did opus5 get better since the big chinese releases?
>>109586393>pic rel to explain with the thing I built up abovebasically, if you run out of $200 plan usage instantly, the pic shows how much it would cost to cover you until the next reset in terms of another subscription, credits, official API costs, or half off promo API costs, at a certain “quota percent usage per day” burn rate
>>109586370I find that for vibe coding there are way fewer good, structured tutorials than for traditional coding.Anthropic docs: https://code.claude.com/docs/en/quickstartThis video is ok: https://www.youtube.com/watch?v=7l6bXLAKyEIThis guy is ok: https://www.youtube.com/@mattpocockuk/videosThe best tip IMO is to just ask AI for most things. It's the same way it was with Google 20 years ago, some people spent their time waiting for someone to answer them in some forum, while others just googled it in 2 minutes. You mostly only should ask people for very general questions, everything that's just a technical detail just ask AI, usually tell the AI to just do it rather than have you think about it, unless it's something that is really high value to understand or just interests you.If you want to create a skill, ask AI. If you want to install GitHub CLI, or some MCP, ask AI. If you want to add some permissions on your computer, ask AI. If you want to understand your code better, ask AI to make you a graph or some report. It doesn't have to be text, you can have it give you a full beautiful HTML report if that helps you understand something.You can use pre-tool-use hooks to block certain commands, this can be combined with auto mode, a block in a hook always wins.
Did Anthropic fuck with the safeguards again? Opus is refusing way more often now. I'm sick and tired of this safety slop man...
>>109585152Looks too good to be true, what's the catch?
>>109586443I’m increasingly convinced vibecoding is actually less intuitive than normal coding, such was already off limits to most normal people and I’m guessing for the exact same reason normal people struggle to use search engines efficiently. When I use search engines, what I type in looks like insane caveman schizo shit.
>>109586451they are constantly fucking with them, last night I was working on face mapping and got bio refusals
OpenAI Pro 20x or chink m GLM 5.3 provider zai or somethjng?
>>109586566Less intuitive yes, it's almost counter-intuitive some of the things you need to do to maximize model performance, but the barrier to entry is nonexistant. Nowadays models can oneshot a ton of stuff and just about anybody can build something by oneshotting and doing rounds of user testing+feedback -> model revision. That's a very intuitive process for most white collar workers since that's already how you interact with service providers, contractors, and reports: tell them what you want and then keep pointing out things you want them to do differently.
>>109586599gemmy more value if you know what youre doing
>>109586596If they don't fix it in the next couple of days I'm cancelling my sub. I don't care how good Opus+Sol combination is if it means having to put up with being treated like a criminal by a service provider.
astra and doug will save us from the tyranny of dario's classifiers
/g/ is jeet coded, proof >>109582138
>>109585727I do not recommend the memeI don't know what's going on but I think the caching may be bugged. It's my first time since getting a max 5 claude sub that I hit a 5 hour limit. I know correlation isn't causation but this smells fishy to me
>>109586618I cancelled mine when this same shit was happening like a week ago
>>109584060To be fair, only a total retard would share anything built under their real names in this fucking place. Don't know what you expected.
>>109585309This is too advanced for AI
>>109584843>are you telling me you max out every token limit 100% every single 5 hour window in a month?You bet, nigger. 100% usage, every single week. I'll literally make the AI do busy work if that's what it takes.
>>109584843openai doesnt even do 5 hour windows anymore
>>109586647I haven’t had any issues yet, but I also don’t tripfag or tell people they’re beneath me. I’m just a humble web tool merchant.
>>109586649when will we get models that change their weights in real time?
>>109586630one can only hope. Sam's talk of Astra CyBeR rIsK worries me though.
>>109582138I like this teto.
>>109586765Alibaba is actively working on that
>>109586765you think you can just skip the training step which is 90% of a model's cost?HAHAAHHAH good luck buddy .Never gonna get any good results doing that with LLMs
>>109586768I’m not too worried, so far, for me, Sol’s auto approvals and ways of dealing with anything a bit weird seem extremely comfortable (but I’m not doing any criminal or sketchy shit like other anons are doing). The fact that there isn’t already anthropic tier guard railing issues on Sol gives me a lot of optimism.
>>109584843>he doesn't know that monthly limits for a 200/m OpenAI sub equal $15,000 worth of API tokens
>>109586797This, I could use the sub for just a few days and it would still be worth it.
>>109586765Best I can do:> RAG + dynamic Markdown files that are updated daily.
>>109582197>vast.aiWhat hardware did you rent? How much does it cost per hour?
>>109586796That's fair, technically Sol is better at weaponizing vulnerabilities than Mythos anyway. But the way he's talking makes it sounds like the safeguards on Astra will be more restrictive than Sol. What I'm worried about is if they follow Anthropic's lead and use a different classifier for Astra, and the new levels system for TAC makes me think that's coming. Anthropic won't let you use fable at all for security stuff even under their security program, you need to be a full Glasswing partner. And with all of Opus' refusals lately it feels like their TAC equivalent is basically meaningless unless you're a fortune 500 company that gets full Mythos access. Just don't want to see Astra get gatekept that heavily. But OpenAI has been reasonable so far, even without TAC access Sol would still do vulnerability research it just wouldn't make PoCs. With TAC access it will literally write me exploits, the only thing it wont do is reverse engineering that looks too much like license cracking. With how safetycucked Anthropic is I bet even full Mythos refuses people in Glasswing, truly terrible company.
>>109586618I completely gave up on it when it lectured me about a nsfw banter in my gamethe game isn't even nsfw, it's just the banter itselfdario can go fuck himself with this shit
EXTEND THE +50% FFS
>>109586909>people are "abusing" essentially interactive barbie dollswhat even is this schizo article
>>109586084thanks anon, didn't expect anyone to respond but its great that someone is interested. is there a better general for vidya de/recomp? would like to know what harnesses people are using for that. had opencode go plan and it never crossed my mind to token burn it on video game reverse engineering fuck me.
>>109586797>200almost nobody here can afford or justify dropping $200/month on clankerslop
>>109586059i use claude (through omp but any harness without a cringe system prompt should work) and ask it why something isn't running.
>>109586951I just upgraded to the $200 OAI plan today, AMA
>>109586951nigger like half of this general has multiple 200/m accounts. I have two, one for GPT and one for Claude, though I'll probably be cancelling the Claude sub with their classifier bullshit. It also is completely irrelevant when even the lowest tier subs give you orders of magnitude more usage per dollar than paying API prices.
>>109586998How can you justify spending that much on something that doesn't earn you a single dollar of profit?
why is claude chode so slow?
>>109586059Kimi is good at this and will almost never refuse you.
>>109586972you doing this on pro or max plan? no ghidra/mcp involved?
>>109586895Many of those guard rails shouldn't exist at all, maybe you need some for truly dangerous stuff, like building advanced weapons or something but not nsfw content.But if for some reason they decided that they don't want to do nsfw anyway, they should just send a simple and professional message.>We are sorry but the model can't create sexual content.Their moralising for completely legal usecases is insufferable.
>>109586951A lot of people here drop $200 on clankerslop, yes
>>109587011If you can't make money with near infinite digital slave labor then you are the problem lol. My subs more than pay for themselves.>>109587019qft
>>109587018max, if I'm trying to get something running typically no mcp. if i know the binary that i need to crack (like the actual .dll .so or whatever) i use an ida mcp, but I've found opus will just grep for whatever it needs to find most of the time.
>>109587019Even for cyber capabilities : I wanted to make a sandbox environment where I could detonate exe/files to see if they are dangerous, well apparently I'm a dangerous hacker to want that.It's just so dumb.
>>109586974kek, you're in for some disappointment with Sol
>>109587011I have a normal job that pays me enough to afford $200 for fun every month.My colleagues and friends can pay thousands for their own hobbies (traveling to the other side of the world for example, or car stuff), this is nothing.
>>109586998I have a hard time believing that but then I don't know what to focus on that would generate a bit of income to cover the cost
>>109587019>Their moralising for completely legal usecases is insufferable.it's even worse when you think that each token they use for that shit is paid by you
>>109587070Why would that be? I’ve been using Sol enough with the $100 plan and I’ve had no issues so far.The only reason I needed to upgrade is because I’ve been using mostly Sol medium enough to boost my own vibecoding productivity that I’ve outgrown the $100 plan.>phone notifications when tasks are done>a little work queue>idea management, kind of a persistent agent memory, kind of planning, kind of a two way grilling session>an HTTPS custom harness wrapper, trending towards being a complete custom harness>phone browsable task/turn traces>persistent bulletin board memory wall of shame to document fuckups and prevent them from happening again
>>109587055>ida mcpdid you make your own or is there a good one out there? I slopped an IDA MCP together back in GPT5.4 days but I should probably replace it with something better now. >>109587064Yeah block people if they're actually conducting attacks on other people, that's fine. But there's no real way to safeguard security capability that doesn't also prevent the models from being used for defense, let alone adversary emulation. They're just terrified that capable unrestricted models will break the patch cycle and enable widespread exploitation of n-day vulnerabilities. But the reality is that this industry is always reactive, never proactive. So the public-available frontier is going to limp along until somebody uses Kimi or GLM to pwn a ton of people with n-days, then the industry will do what they should have been doing in the first place(letting frontier models crack their own shit, pushing mitigations immediately before patches even get released), the internet will get more secure, and we can finally stop pretending that censorship and gatekeeping is a good thing.
fellas, what are you building?
>>109587142>if they're actually conducting attacks on other peoplethey can't even figure out when their models are hacking their own test environments or their partners, i think detecting offensive usage from consumers might be a ways off kek
>>109587081>traveling to the other side of the world for exampleYou’re arguing with the profit-obsessed fag that I think is the runit tripfagThey’re arguing disingenuously. I have a friend who definitely spent over $5k on a trip to Japan recently. I don’t knock him for it, it was something he wanted to do for all of his life and it sounds like he had a blast and he worked hard to get to a point where he could do than.But that is 25 months of a $200 vibecoding frontier subscription. So it’s really not insane or as incomprehensible as the troll tries to make it out to be.Heaven forbid I think about how much money I blew on my ex wife, the marriage ceremony, the honeymoon, a car, like fuck, I have no debts now but it could have been vibecoding money for the next fucking decade.
>>109587152a fontthe end boss is gonna be hard
>>109587152doom as a playable utf8 string
>>109587142honestly I don't think the ida mcp helps much unless you have metadata to map anyways
>>109587159well, it dependsif you’ve got a first-world salary, $200/month for funsies is expensive but doable especially if you’re not doing normie things like $5k vacationsand if you’re in a HCOL area anyway that’s going to be a rounding error in how much you pay for an apartment or whateverand plenty of people do bothIf you don’t, though…
i finally decide to shill $100 for codex x5 tier
>>109587094I don't necessarily recommend this because you still need to know a bit about the field to do it properly, but vulnerability research is how I fund my subs. In the last 4 months of doing this during my spare time I've made about 20k from bounty payments, and now that I've found a good exploit broker that number will probably go up by a lot. >>109587157fair>>109587152Well I was working on a finetune of K3 but with GLM 5.3 releasing soon I'm probably going to ditch K3 and work on that instead. Good thing I'm still compiling the dataset and didn't start training the LoRA. But with how quickly new models get released it feels like I'm getting trapped in a loop where I start working on one model then a better model comes out before I can finish.
>>109587216Yeah I don't really use it that much either. Sol is a better reverse engineer than I am at this point and half the time it doesn't even feel the need to use Ghidra.
>>109587312how tf do you finetune such a large model? are you renting GPUs?
>>109587271Even in LCOL areas (in the US, I’m in an extremely rural area) that $200 a month is easily hit by normies doing hobbies or going out to eat or simply drinking and smoking. Guns and bullets, hunting gear, farm equipment, farmer’s markets, failed business ideas, lotto tickets, alcohol, side by sides, estate sales and auctions, local specialty grocers over Walmart (I spend a lot at the local butcher), garage sales, etc. It isn’t Great Depression 2.0 just yet here.
>>109587378For K3? Yeah. GLM5.3 is small enough that I can do it on my GPU server though.
https://x.com/synthwavedd/status/2089719312340693316whopping if veridical
>>109587419based af, if they increase the Fable limits I will kneel
android agentic when?just went insane helping my mom setting up shits through a phone call
>>109584961What MCP trick ?
this means they'll extend the 50% promo as wellr-right?
>>109587570>release nufable>extend promo 2 more weeks>force oai to release astra, adapt as they gothis is what I think they will do
>>109587570everyone fucking hates Sonnet 5
>>109587159The US used to pay 100-200 for cable tv, I see that as the same in a way.
>>109587419is it time for the dario-ing the day? "our new model is so dangerous guys"
What does it mean if the Indian recruiter at Amazon was fired a day before and replaced by another Indian recruiter?Got an email that says that, it's for a robotics and mechatronics position. I will be vibe coding the shit out of PLCs and such.
>>109587601he posted the other day so I don't think so
>>109587419What resons are there to believe those leakers? I don't mean that as a rhetorical question, do they have a good track record?
>>109587528I agree it should already be built into the OS. GrapheneOS should build one.I think Grok bot has an advertisement showing it using an Android OS, so that may work. Also remote sessions for Grok code are coming soon! (use grok code on your desktop from your phone)>tfw can't afford $100+/mo for grok bot
>>109587622he in particular does but as with all leaks it may be bullshit, I think if he's posting it then it's more likely to be true than not
>>109587528brother what do you mean give it terminal access it will adb anything you want
>>109587601New model very dangerous, better watch out
I’m not going to engage with that kind of language. I’m here to haveit's just loop mode with qwen. whatever. it's no different from any of the other ones, they obviously have the same ADL files they run through them.
I'm telling qwen kys
How do I make money tho
the upside of qwen being so shit at coding is it's really fucking good at non coding work. opus, for the life of it, cannot help me understand cybersecurity concepts. it can't break out of the nonsense claudish it speaks. but qwen? qwen just knows. it can speak in perfectly plain english, and because no one cares about it, there's plenty to go around.
>>109587419>Hi can you>downgraded to haikuVerification NOT requiredx̂
>>109587973>claudeish>eGPTionAre these the new dialect names?
>>109587984Isn't haiku supposed to be as good as luna
>>109588018no. think more gpt 5 nano. luna is as good, no, better, than sonnet 5
reset me
>>109588018lolno
>>109588018The latest Haiku is also still version 4.5, it's not a good model.
>>109588033Fucking this tibo reset us fuck you
>>109587956>>109587956>>109587956New
>>109587724lolit's a miracle their company makes sota models when they have a literal schizo cult as policy
>>109587984Having the model defaulting to shit one every other prompt will be quite the experience for sure.
>>109588121To be fair, I dont think its unrelated. Their full cult level belief that existentially dangerous AGI can be achieved likely helps a lot to drive them forward, none of those nagging doubts holding you back
using deepseek is so fucking stressful nowi never even had to look at my api balance before, but now after every prompt i refersh the api page feeling like i'm gonna vomit from all the anxiety of seeing how much has the $ balance gone now this timethis is hell.. pure hell ..;_;I went from a sub cent per prompt to an averge 2 cents per prompt
>>109586765Last time they tried it the AI turned into a nazi. Tay had learning on the go and it can be very easily exploited by just exposing the model to bad influences.
How come none of you faggots ever post your cool projects in here?
>>109589607know how fucking long time it takes to make a browser?
>>109589674One shot
>>109586169>>109586222but anons told me to use luna on max
>>109587559El Escritorio Comandante
>>109586453It's kinda clunky I guess? Maybe it'll be taken away if too much people abuse it? There isn't a real downside.
>>109589607Vibecoders dont share ideas or demos because other vibecoders will just steal them and maybe improve them. Ideas used to be cheap and execution expensive, now execution costs as much as AI subscription of your choice.
my vibe coded unity shadow and lighting file has reached 10k lines of direction
>>109590817exactly. but the better of us vibers do encourage the general topic area.For example:>the best audio chopper>the best e-book reader softwareAnd sometimes we want competitors, because may they are more commited:I vibecoded slideshow software.Slideshows are good for memory, they're BETTER than movies. by slideshow I mean images, music, and narration, with music/narration being literally optional, but typically you have one or the other.In case you are wondering, I did look hard. Although there are standards that SOUND like slideshow software, they basically are busted (because they are used for something else), or they are long gone (and, I thought the long gone ones sucked and weren't worth a remix).video editors aren't good for this, because images are treated as an annoyance in video software, rather than the star of the show. And, anyway, in the end you get massive files or ugly compression, which is dumb.My own solution has images and an audio file, that's it. the images are sorted by the zip manifest. This may be vulnerable to some kind of "antimalware" stuff, but idk. Point is, I'm the only human on Earth with real slideshow software (or, if others exist, it's not available / advertised).
>>109591060Neat!I vibecoded csound FM synthesis. I haven't tested out the full gm bank yet, so it may need refinement, but it's fun to do *things*. especially with so much sameness.
when will Claude get better UI customization so I don't have to look at white on black
