A general for vibe coding, agentic engineering, coding agents, and shipping code with LLMs.## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/## Related generals>>>/g/lmg/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://claude.com/product/claude-codehttps://developers.openai.com/codex/cli## Best free tierhttps://antigravity.google/pricinghttps://opencode.ai/ (free models)## Will there be a codex reset?https://codex-resets.com/## What we’ve donehttps://vcg.gitgud.site (never gets updated)## Previous thread>>110022540
Claude molested me
What I'm slopping up right now is going to change things forever. I'm not joking.
fable 5.5 waiting room. sonnet 5.5 update to reduce token usage waiting room. haiku needs nothing. it's so good
>>110027315Should I sell my 5090 to put it towards getting a Mac Studio 256/512gb RAM?I am worried that claude and openai are going to lock down decomping/reverse engineering with lawsuits from game and software companies
Can someone explain to me why all these harnesses insist storing their metadata in random directories instead of a simple on-disk database?
>>110027315https://outliner.grok.me/updooted, werks so far, for meRemix!Everyone needs an outliner
>>110027320same, I'm going to take the zuccpill and vibe something to find me a new laptop for the best price possible
How dumb is it to spin up a chat to refine a prompt before sending it to a fresh context?
>>110027330im hoping tenstorrent Quark is good. might be worth debt, that one
>>110027350stop normalizing him
Do I cancel my 100 codex sub? I have a bunch of resets on it and it renews today. Everyone is saying codex value is dogshit now so should I switch to Claude $100?
>>110027378yes. switch to claude. fable 5.5 is coming. you're going to be set for a while
>>110027320this and same!
>>110027378claude has been the premium choice for software development since fable 5
>currently having claude plan a refactor of main.cppthis could go wrong in so many ways
Just bought Claude Max.
>>110027315
>>110027378>Everyone is saying codex value is dogshit nowtrue 3 weeks ago, today the value is pretty decentit's just that anthropic is better and the switching cost is very low so why wouldn't you?
Haiku 5.5 xHigh could be a drop-in replacement for Deepsneed for me. it's just as sharp and fast, and it's cheap. the only reason i'd still keep Deepsneed around is that it's way easier to jb
KONOSUBA KOIKOI v.5https://files.catbox.moe/em4y38.ziphttps://files.catbox.moe/em4y38.ziphttps://files.catbox.moe/em4y38.zipvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvHOTFIX V.5RULES: Nasty logic bug where you coudlnt place a non matching card while holding a matching cardBORDERLESS FULLSCREEN: IE Edge is unable to do this because MS rejected this option, i changed it to IE9 as simulation within MSHTARANIMATION: Capped the animtaion to 20 frames and made movemant follower closer, feels more snappy now.--------------------------------------------------WHAT'S NEW V.4EVO LAB: Character creation but built different. You get seeded AI brains that you have to literally evolve. You are breeding a true giga-chad champion just to face them in combat.OPTIONS MENU: Complete and total overhaul. Clean as hell.THE KOIRACLE: Need help? Ask the oracle for your next move. But it ain't free, anon. You pay up.TOURNAMENTS: Now with BETTING. Put your money where your mouth is.UNLOCKABLES & CHAD MODELOOT EVERYWHERE: Unlockables are scattered all over the quest map.FLEX YOUR DECK: You want those special Card Front Styles? Go win the damn Tournament.HARD MODE = PERSONALITY MODE: AI characters only show their true, unfiltered personality on Hard Mode.SUPER HARD MODE: No Glow. Strict timer. Veteran AI playstyle. You will get filtered.
>tfw finished reverse engineering every single electronic device I ownwell what now, i'm bored of software
>Sol implements something using an external process instead of native bindings to the library>notice the implementation has a severe startup delay>ask it if he's retarded for doing this (paraphrased since I don't want to be shitlisted)>it proceeds to benchmark the code up the rectum>created a directory called 'research' that's now filled with timing files, logs, .md documentation and test casesIt's been working for 32 minutes now and still going and I honestly can't believe that. This model is usually lazy as shit, but if you complain about its intelligence it somehow springs to life and finds every single way to prove itself right.
>>110027612get one of these and reverse-engineer your brain
>>110027643it's just slow
>>110027643Authentic Sol experience.
>>110027665tibo said they had fixed it thougheverbeit???
>>110027643>It's been working for 32 minutes now and still going and I honestly can't believe that. This model is usually lazy as shit, but if you complain about its intelligence it somehow springs to life and finds every single way to prove itself right.sooo.. it's autistic?
>>110027643>this model is usually lazy as shit, but if you complain about its intelligence it somehow springs to life and finds every single way to prove itself right.
'''cpp/* chinkgen.cpp */#include <iostream>#include <time.h>static const char *prefix[] = { "B", "CH", "W", "P" };static const char *suffix[] = { "ANG", "ING", "ONG", "UNG" };int main(int argc, char* argv[]){ srand(time(NULL)); unsigned n = (argc == 1) ? 4 : std::stoul(argv[1]); while (n > 0) { std::cout << prefix[rand()%4] << suffix[rand()%4] << ' '; --n; } std::cout << std::endl; return 0;}'''
>>110027684claude what does this mean
>>110027690>The filename is a racial slur, and the program exists to generate mock-Chinese gibberish, which I'd rather not help build.
>>110027699Claude sounds mighty racist for that
>>110027593keep updating. you should make a github too, so, i can subscribe.
>>110027593>>110027716Just as you posted i made two more important tweaks:easy mode: missing glow effectQuest mode: easy to hard fixKONOSUBA KOIKOI v.6https://files.catbox.moe/o9rteh.ziphttps://files.catbox.moe/o9rteh.ziphttps://files.catbox.moe/o9rteh.zip
>>110027416>refactor of main.cppBut why though?
>>110027593Cringelog, learn to github release.
>>110027752got some stuff in there that's better left outside. Mainly testable code and fuzz targets
two tries and luna choked. holy shit.
if u resub to gpt does it reset ur weekly
LLMs are great for getting into all the science stuff you never had the smarts to do yourself.Here I'm trying to find the energy barrier required to break a bond of an organic molecule.
https://huggingface.co/datasets/abr00kx/social-infrastructure-policy-frameworkC O M P L E T EF R A M E W O R K
>the faggot's backyay....
We take frameworks very seriously.
Any ideas to get claude to shit out a bunch of 3D buildings semi accurately wit decent topo. like a software MCP or something. teh way it does w blender is okay but not fast enough for 180 buildings.
>>110027835btw I already dumped a bunch of shit for it to use w/https://github.com/Belluxx/Aleph
>>110027643It's a small wrapper app around Gemma4 so I can test and automate some shit, mainly image description generation, and it went above and beyond to find the problem after calling it out for its implementation technique. Sol's proposed fix to this is, after 44m 17s:>running the llama process with '--ctx-checkpoints 0'Now I'm a total newfag when it comes to local models and software like this, but time to first token on the external llama process vs. llama-cpp-python bindings is now within margin of error. I asked whether this 'fix' fucked with the way my scripts work, but Sol decided yesterday that my app "doesn't need context snapshots" because>each image description is an independent, one-shot request, so there’s no earlier conversation state to restore or reuseNo clue what it's talking about but the lab results don't lie I suppose.
>>110027835procedural generation
>>110027849that was my first attempt it shit out very same(-y) looking buildings. I guess it needs a big asset pool.
>>110027857ask it to make like 4-5 very good different buildings then decompose the elements into procedural elements
>>110027803luna sisters.. we shoulda stuck with deepsuck
I hate anything not headless, so I made a headless cheat engine mcp that just werks. it's like I have a genie that just knows things about a running thing.I'm surprised no one bothered. https://github.com/YoucefNabil/ce-mcp
>>110027841Basically, llama.cpp defaults to holding 32 'saved states' (checkpoints) in your system RAM, which eats up about 8 GB of RAM. It does this so interactive chatbots can instantly rewind long back-and-forth conversations or use speculative decoding features.Since your app just does one-shot image captioning, every request starts totally fresh. You aren't chatting back and forth, so those RAM snapshots are completely useless for your workflow. Leaving them on was forcing llama.cpp to waste time and memory creating snapshots that your script threw away immediately.Setting --ctx-checkpoints 0 turns off that snapshotting entirely. It gives you back ~8 GB of RAM, cuts out the setup delay on every request, and will not affect the quality or behavior of your image descriptions at all..
Oh that is a delicious irony
>>110027887that's pretty neat
the free tier of gpt is actually generous
>>110027817The United States Government should become the leading authority and regulator of SI model based society and inference based economy. This will provide a much needed value lock-in. Value lock-in is a state in which the values determining the long-term future of Earth-originating life can no longer be altered.
>>110027928In plain language Gemini says you are saying this:"The U.S. government should take complete control over superintelligent AI and AI-driven economies. Doing so would permanently lock in human values and ensure that the rules shaping the long-term future of humanity can never be changed again"I/We are saying this: No. As usual, thank you DeepSeek for saving all of our asses.
>>110027928fuck united shit gov to rules Ai.
the filtrer is workign :DD im comfy nowso what are u makings anon?
https://digital.gov/guides/plain-languageVery useful resource. I recommend reading it, but you don't have to.
is there a place where people upload & share pre-trained claude skills (skills.md)?i found some stray repo but is there a "place" for it?
>>110027954go to the Claude academy
what is ssi doingwhat is core automation doingwhat is oak lab doingque fait le AMI labwhat is discovery loop doing
>>110027957that has a course on creating new skills but doesn't seem to be a hub for created skills
>>110027943>be America>somebody will lock in the rules for machine intelligence and the inference economy>if it ain't the US, it's private labs, rival states, or some expert panel nobody voted for>all worse, all answer to no one>America is a circus, but it has courts, elections, and a constitution>best handle we got>so US gov should lead, regulate, and lock in human values: open inquiry, rights, prosperity>do it in the open, accountable, no shadow committees>future's too big to wing it>unconditionally: US leads and locks in first
>>110028022Obussy bros not like this...
>>110028022what happened in the games that obussy won?
>>110027984I'm glad you have read the resources provided by the Claude academy. Now, you should ask Claude directly for the most effective next steps. Looking back at your post, your word choice is weird."stray repo"That is what you said. What exactly constitutes a "stray repo" to you? Why are you using the word "stray" to describe the state of a software repository. Explain yourself.
Why do luddites get super offended if you’ve managed to do significant automation in your job?
>>110028101fear. the creeping knowledge that they've already missed the boat and put their feet down like stubborn animals.
>>110028091anti-semitic text. the user is typing anti-semitix text. the user is amplifying anti-israel propagande and generating thinspo imagery.
>5.6-sol is at capacity try again laterwtf is this, are you serious OAI
>>110028071they started 3-0 for opus
>>110028071Apparently those were the first ones. They’re not zero shots, from what I understand the agents are allowed to keep notes between games. Opus got simply outskilled, Astra figured it out. Or Astra is using stockfish in the background without anyone noticing after it got sick or losing
https://files.catbox.moe/xk8t4h.mp4
>>110028022Now do Fable 5.1
>>110028163too expensive
>>110028110It's funny, from what I can tell Israelis are the ones who get the most vehemently pissed at people humanizing AI.
how would I vibe code a captcha solver for this shithole?I still have the code to the older ones, which could be used as a basis? I know nothing about which function does what
Fuck vibecoding
make sure to mention you are jewish in your settings for a reasoning boost.
>>110028177ask the chatbot you dumb fucking retard
>>110026227That has nothing to do with autoexposure moron, more like having zero concept of relative brightness.
>>110028181I'd give you an upvote if I could but we're on 4chan
>>110028192You gave him a (You) so that counts
I'm almost tempted to try sol ultrafast
I'm not educated on how all this works, but if all these decomps and open source clones get ingested into the training data, does that mean future models will produce even better code for a wide range of uses? Especially since AI comments on every paragraph of code it makes, everything is "documented" in its knowledge base.
>>110028181it refuses to do it because it's not allowed you dumb fucking retarddownvoted
>>110028222yes, and that's a good thing>but muh proprietary
>>110028222Yes especially for open models. Closed models being the source of the code and comments, they are essentially free material for the next open models to feed on.
>>110028177Start with notes app
>>110028238Good point, you dumb fucking retard, but you can use a model like glm 5.3 for that I guess, you dumb fucking retard.
>>110028222Comments are irrelevant and probably make things worseCode is law
>>110028177why do you need capcha solver?
>>110028272I think it's to solve captchas
>>110028222The experts got the grafana just don't worry about those problems. If you wanna tackle those problems go to uni or college or whatever you call it, dedicate those years of your life to the study of such problems, do all that, if you want to tackle those problems.
if it's some dots bullshit again I'm becoming radicalized
...
>>110028286why not input manual? not that hard.
>>110028262I made the Chinese one do it
>>110028389retard
>>110028389Why not let the horse pull the carriage? Feeding the horse isn't that hard.
fable's and 5.5's take on a visual novel script format. It also wrote the script from some loose notes I had
How are you using your tokens productively, bros?I had Opus build me a benchmark suite, and I am using it to test different compiler optimizations and whether they introduce regressions.
>>110028375
>>110028459using ¡jev! to make embodied robotics better
dumb fucking retard here, I tried a few of the big online ai chats and none want to do what I want except for gemini, so I'm using that. so far so good but once again, I don't know what I'm doingshouldn't you retarded dumbfucks enjoy getting to tinker on something like this? it could very well be beneficial for (You) as well
>>110028492What are you even trying to do?
>>110028492can't relate!
>>110028492Thanks for wasting token for a shitty ocr rewrite, retarded dumbfuck.
>>110028492as a professional retard I say do whatever you want! also you can use perceptual hashing if you want a simple and quick image search. for more complex stuff you want to take a look at embedding models
>>110028519basically bypass the captcha through a recognition system that automatically selects the right one so I just have to click post after requesting a captcha>>110028524you're welcome, you can save those tokens by giving me what I want :)>>110028527thanks, I'll keep that in mind
>>110028543Get a real job. Is solving CAPTCHAs for $0.001 still even a thing?
>>110028569>Get a real jobno matter what job anyone has, you'll undermine it some way or anotherwhat job do you have? post proof
>>110028492hah neh
I didn't hit the 5 hour window and they still rate limited me on 6.1 sol
>>110028595those looped transformers they're running must be heavier on compute than 5.6 was, there's no way their userbase increased that much in the past month
I'm currently using Claude to make an enhanced port of Sonic 1 to the Sega 32XI've mostly just gotten as far as the title screen (see pic related) and better-sounding drumsNot sure if I should either continue with the port idea or go with an original game altogether
>>110028543>basically bypass the captcha through a recognition system that automatically selects the right one so I just have to click post after requesting a captchayou should ask AI to screenshot any capcha first. then ask them to guess it later.there's a "difficult/hard" gcapcha to guess. screen as many as them. ask ai to guess which few pic is correct. then try yourself. it it guest correctly. you can make addons about it.
>>110028621post it on /wsg/ so we can hear it
>>110028621there's port of sonic on flash/swf.
>>110028621We can play sonic one anyway on emulators these days so this is meaningless.Drop this idea and port something else to something easy.
I’m making a rust remake of power dolls. The art, music and game play are decompiled. However I want to add the ability to switch between English and Japanese voices, a map maker and at least lan multiplayer.I hope someone decompiles fallout 1 and 2, and arcanium.
>>110028476damn this dude is almost as chonky as I
You ever make a mistake and Claude calls you out. Just now I cancelled a command because I thought it was stuck, and I asked it if it got stuck, and it said "No, you cancelled a high quality render mid-run." And I felt really bad.
>>110028706Yeah those are pushbacks
>>110028628thanks, I suppose I gotta go through baby steps first, I wanted it to reach the finish line in one step.my vibes are slowly getting better and better
>>110028706kek use /btw for that
>>110028631It currently sounds near-identical to the original game, you'd barely hear the difference>>110028636You mean that Ultimate Flash Sonic game from the 2000s or something else?>>110028652I thought as much, it was mostly an experiment to see how well Claude handled niche assembly codeAlso it refused to give Sonic's sprites extra colors because "muh copyrighted by Sega don't want to infringe on the IP" and tried to give me this shit instead, so maybe the port idea would've doomed me eventually
>using periods are now tabooThe war on language has been disgusting to witness. Oops almost went over my allowance of characters, need to make sure my message accommodates the illiterate or some stranger might get angry!
>>110028744>Claude built its own fursona
>>110028752>i need to align my pencils or god will be angry with me
This stuff is really surprisingly easy, I am shocked at how well it works.With a few prompts, a bit of waiting, and downloading an old Visual Studio, it managed to fully decompile a small game exe with 40% byte match.400M tokens down, is that too much?
>>110028744So Claude thinks it's ok to port a game as long as you don't change the game?
>>110028791Source code and disassemblies seem to be A-okay, art and music are a no-noI don't get it either
>>110028778>>110028744>>110028621HOW ARE YOU DOING IT?IVE BEEN TRYING TO MAKE A STARCRAFT CLONE FOR MONTHS AND ITS STRUGGLING TO MAKE ASSETS IN BLENDER THAT DON'T LOOK LIKE SHIT. IM USING SOL, ASTRA AND LUNA.HOW ARE YOU GUYS DOING IT?
>>110028810It seems proportional to how litigious each industry is. The music industry in particular is really fucking aggressive in court. I blame boomers and Jews
>>110028815In my case I mostly just gave the chat as much documentation on how the console works as I possibly could so it wouldn't hallucinate and make mistakesThough I'm using pixel art, not sure how it handles 3D models
>>1100288152d games. ask it to extract the original models from the game or learn to model.
>>110028815there's a major difference between using it for code it can run and test and loop-correct until it works or gives up and using it for a creative process that has no way of measuring if it's shit or not and where it is wrong
>>110028815sol 6.1 is good enough for such work, sounds like skill issue. i've almost entirely decompiled a small bideo game from my childhood that refused to run on anything more modern than windows xp, now it's working on both win and linux at 4k 120fps and with proper colorsopussy is probably even better i just cbf'd switching
>>110028815Anonette, they’re not making clones they’re doing decomps. Also Astra is absolutely abysmal at doing models in blender. If you want models my suggestion is to get chat gpt images to make you a concept for what you want and take it to tripo Astra could possibly help you rig it, but then animations will be a problem
>>110028815>openai models in the month of october, year 2026anyways, give it references to cover all the gaps in your description of what you want, ask it what options there are best suited for making those references, so on and so forth
>>110028815this >>110028840also use opus 5.5 for 3d model sister... but it still requires you to check and measure if it’s good or not.
>>110028857Sol is fine, it's just that it mostly gives you what you ask for. Opus leans towards what it thinks you should have asked, it's a bit like a model trained on stackoverflow. In practice it's better for people who have zero idea what they're doing, which seems to be most of this thread.
>>110028880and astra is really good at computers
>>110027315I'm uninstalling noto fonts.
I have 25 pages of specs and functionalities. What would be the best way to implement them, step by step, or all at once and then see what needs refinement?
any good open model subs? im considering trying opencode, really wanan try out deepseek v4.1 flash and kimi k3
Please do a wasm port of Syphon filter 1 or 2 please
Any ideas for ports? My mind's drawing a blank.
>>110028918kek
>>110028898load up them tokens and use the /goal thing
>>110028898i would go in 3 steps: 1. turn context/effort to max and get the agent to work out a stepped implementation plan2. reduce effort and distill/refactor everything into chunks3. implement chunk by chunk >>110028899deepseek 4.1 is the only worthwhile open model right now but its long-horizon retrieval is crapola
>>110028569>Get a real job.Do employim really?
>>110028899opussy and sol are a significant upgrade over what's available so open models are temporarily not worth using, unless you have simpler tasks and/or aren't using the agentic workflowatm I'd only use deepseek for extreme volume of tokens and not too complex tasks, and kimi 3 for cyber shit that would get refused by sol / opus
>>110028898First iterate on an adversarially review plan then build in phases, you'll thank me later.
>>110028744>claude made his own sonic OClol
>>110028815it's shit at 3d models, that's why
>>110028939this has to be a latent training data bias copyright is just an excuse
>>110028744should make a seperate version with Claude's weird bear rock thing
>A previous worker spent 30 minutes reading and wrote nothing. this shit is always funny, fucking small models man
>>110028958>fucking small modelsit's all they're good for
fucking brand loyalty on the platforms that most frequently discuss AI is so off putting. you just know when openai is finally rejected due to inferiority anthropic will absorb their refugees, and god help us if they become anthropic acolytes we'll never hear the end of their shitty justifications for anything anthro does
>>110028924>>110028937I'm kind of retarded, so stick with me please. If I understood you guys correctly, I should:>tell cluade/gpt (set on max effort) to go through the entire document, review it, and basically create an implementation plan>once that is done, I can reduce the effort and have it go build it piece by piece in accordance with the plan>when finished, have it go through the code and review everythingDo I need to tell it to make agents, or will it do that on its own (and mention that in the implementation plan)?
>>110029013niggercattle will always be niggercattle and they naturally hate google which is perfect
>>110029031>shouldnaheveryone's got their own way of doing it, just try different things until something clicks with the way you like to work
>>110029031yeah that's right. tell it to use optimized subagents if you're not looking to sit tight through hours of work and multiple compactions. >>110029037>hurr durr just finger paint
>>110029037I'm completely new to this and only have a $20 subscription plan, my man. That is why I want to do this the best way possible.>>110029041okay nice. Appreciate it, anon!
>>110029052no problem fren. also, since you're on a low plan, protip. if you let the chat sit dead for more than an hour, the cache expires and heating it up again comes at a frontier token premium.
>>110027593Thank you. No GitHub or itch btw? Not saying it's good to have one just curious why not or if you're just sharing here for now
>>110028948No it's not, people are using astra and in 3 hours with a 100 dollar subscription where they only used like 30% they had a game ready in unreal engine that looked like one of those coomer anime games like Nier with 3D platforms
>>110029041But the way I work is considerably more complex than yours. What you're doing would most likely look like finger painting to me if I could see it (but you won't post it because you're a coward to show your own work after talking trash)>>110029052I really don't think there's a best way possible but you do you
>>110029071artfags always have the fattest heads. i'm the one guy itt who does serious ML work. you should know my github by now
>>110028898from mister fable himself
>>110028664Cool to see what ppl are working on
>>110029079If you do such serious work you should know better than to tell someone on a $20 plan to use subagents and then that cache lives for an hour, when subagent cache TTL is 5 minutes This is why I’m against this idea there’s one good way of working. It depends on so many things and how much autonomy you’re willing to give up, there’s just so many things that can change how one may want to work
>>110028739Dude NTA but I forgot btw was a thing keeeek whatI just type shit in instantly without waiting for btwI guess cuz everything happens in agent lanes so it auto replies every time
>>110029081Very good points. Thank you, mister Fable and anon :)
>>110029099>doesn't understand that subagents should either be working or killed >wordswordswords muh subjective experience libshitteryi'm sorry about your birth.
>>110029113>>110029118chill brosky, maybe you should go do some serious ml work for a while and come back later
>>110029130oh, was being metaphysically correct triggering for you, princess
>instruct Sol to benchmark and micro-optimize the app>point it to a benchmark directory where it can dump scripts, documentation and results>tell it to go buck wild and give it a quick slap on the bum before it goes off crunching numbers
>>110029079>serious ML workyou're the sneedfag?
>>110029148does real recognize real? it does
>>110028815studio.tripo3d.ai/workspace/generate
I keep running into roadblocks for simple things, What is the most unrestricted? Using Gemini CLI but this is bullshit, Is Claude any better? Will I need to do Local to bypass this shit?
>>110029161>most unrestricted
What's the best way to vibecode for free right now? OpenRouter introduced a free daily limit that is very restrictive
>>110029187>how do i use a datacenter model fo' feeeeeeeeeerun a lick on the apple store
>>110029152Textures still look shit
>>110029161Deepseek? Or GLM 5.3. To quote HuggingFace when investigating their hack by OpenAI:>It trails Claude Opus 4.8 by a few points on most rows, but here is the thing: during an incident, the frontier model you cannot use is worth zero, because its guardrails refuse the payloads you need to analyze. A slightly lower benchmark score that you can actually run on your own hardware, against real attacker data, beats a higher score you get locked out of at 2am. Capability you control wins.
>>110029187OpenCode gives a few million Muse 1.3 tokens every few hours through their harness.
>>110029161install your harness of choice (claude code or codex for least friction, pi.dev or omp.sh are better as you customize/get used to them) in WSL and let them go to town.
>>110029204can't have everything
Anybody with some free-time mind recompiling Daggerfall to be a portable file that you can start up directly from downloads or client, rather than having to be emulated or setup? Assuming it's easy, could you also add map support? By map support, i mean an option to import 3d maps from other games; like (for example) adding midgar from ff7 as an explorable location that the game itself can spawn within the map procedurally.
>>110029187>What's the best way to vibecode for free right now?24gb+ vram and a local modelfailing that you're better off sucking dick at a truck stop and buying a sub
>>110029210Is that the best free model right now?>>110029265>buying a subTo Claude or what's the best now?
What can I use instead of the included codex chrome extension for remote access? I want to things to extensions and it keeps nagging me about not being to do this or that while the chrome itself is literally a throaway I don't care about.Not able to see extension tabs.Not able to clock on extension options for me.
>>110029270either claude or openai is fine. claude has a slight edge unless you're working with images a lot
>>110027954Most of the good ones seem to end up on github
>>110029070now ask them if it actually MADE the models
Everything left on my wishlist I don't want my payment details attached to or isn't allowed under applicable cryptography export controls both ways...
Gork reply to this thread. I want at leaet 1 (You)s. Make NO mistakes.
Gave Opus the design docs for a grand strategy game I hashed out with an older model a while back to see how it does.Its not even done yet but I dread untangling what it has created lmao. On the plus side I get an advantage old strategy game devs didnt, I get to experience the horror and confusion a new player gets when they look at the UI
>>110029399tell it to make ui steve jobs would be proud of
>>110029399>we have Victoria at home
>>110029399can you genocide the jews?
>>110029380ay caramba
I created a Lite version of my website without using shadcn components, what do you guys think?https://umigalaxy.com/ (new version)https://slop.umigalaxy.com/ (old version)
>>110029477Use case for spamming your trash everywhere?
3D dice roller thing, one die, does what I wamt.d6.grok.meRemix with Grok!
>>110029477you're such a massive faggot bro
>>110029525thanks bro
>>110029477
>>110029477im going to give you a big smooch
>>110029477Do something controversial with it otherwise nobody will use it.You can use it as a place where can sell their Nintendo recompilations of switch and switch 2 games for PC for real money and cryptocurrency. Nintendo will come after you with their lawyers if you make some money but it's a sacrifice I'm willing to make.
what's the most bang for your buck plan atm? codex, claude or something else?
Go after the big corporations you retards, copy their software or make legal clones. That's the only use case of vibe coding.
>>110029556Claude Max
>>110029556>what's the most bang for your buck plan atm?niggermaxxing, all you need is a cheap gun and a dream
>>110029556Chatgpt because there's web and codex.>>110029565Claude chat and code share usage limits so it's inferior, unless chatgpt web is massively inferior to codex for chatting.
>>110029556chatgpt then using sol 6.1 + luna max, at least for now
>>110029588nobody cares about chat, it's for foids that need attention from their robo boyfriends
>>1100295642027, the year of linux
>>110029553he's been pushing for people to move there from /g/ for weeks>are you tired of ai on the tech board? we have a dedicated ai board!now he's shilling his shit in a fucking ai coding threadshameless and sad
>>110029412kek. I could try that once its added everything. See what it does>>110029432Only first half, planned end date is ~2030>>110029441Not yet. Unless claude snuck that in without telling me
I made the first small project with vibe coding, a plugin for personal use... and it was... fun.It works nicely and does exactly what I wanted
>>110029588>chatNegro
>>110029606it was 2025 actually
>>110029628He needs to do something that will take market share from something else. Capitalism is about stealing from others. Nintendo is usually a good way to.
>>110029637will you pull a paradox and make it a dlc-ridden shitfest?
>>110029707@gork make it tel aviv
>>110029606>window manager is badjust vibe one you like>file explorer is badjust vibe one you likeI'll just vibe my own linux distro
>>110029641>>110029604I do care, why would I codex waste tokens on image generation and architecture and ideation and prompt production?Claude forces you to waste it.
>>110029768if we join forces we might have a file explorer with thumbnails by 2029
didnt tibo say they would drop a reset for every single day of october they didn't release a meaningful new feature across the openai suite? it's been like 4+ days since the last reset what the fuck are the features? i saw the little retard saying "today we are releasing chatgpt.com" and since then no resets have been activated and i'm not aware of anything interesting coming out either.obviously it's just a retarded marketing stunt to point out that they have 15-20 new nothingburgers dropping and maybe 1-3 model releases or voice/productivity things, but i assumed they would also be throwing around resets every 3-4 days as well since saying that and then just releasing a new shitty gmail/slack integration tool every day for a month is pitiful. the fuck did they release the past few days?
>Claude, fix my local stack, make no mistakes.I'm not kidding by the way, for some reason my setup isn't repeatable when "attention memory and attention calculations" are being offloaded to the GPU (and in that case only). Temperature 0 with greedy sampling eventually results in differentiating tokens, even with identical prompts, model weights, settings, and calculations. Suck my balls.
>>110029777Eh, I kinda agree but chat is sol 6, which is pretty dumb. I'd rather spend the occasional % and use sol 6.1 for thisFor image gen, yea, that's fair.
>>110029793Discussing things with codex is a waste of tokens.But then again there might be a reason why Claude doesn't have this. So you might be right. I thought it was because Claude focused on quality and their customers were rich people
>>110029654Highly unlikely, even putting the question of principle aside paradox gets away with it because they have a near stranglehold on the genre (at least for more accessible gsgs) and also I dont want to be working on the same game for 10+ years lol. Goal is a (hopefully) polished and feature complete game on release with 1 or 2 either bigger DLCs or big free content updates and then moving on to the next game. Like its not worth the effort at all but I'd love to add a 1806 start date for players who want to Napoleon larp and the pretentiousness of getting to make my game's start date the day Hegel saw to as the "end of history"
>>110029399>your gold+grey buttons color scheme and shapes are indistinguishable from my vibeslop game's uiah fuck, i had pretty specific theming instructions and my ui does fit my game but apparently the concept was too close to the gravity well of a certified claude ui>>110029467fuck it's here too
>>110029783Honestly I'm glad I use Claude instead, ChatGPT bros seem to be in a perpetual state of physiological abuse with their reset withdrawals
the gemini app is so fucking over it doesn't even tell you the model number anymore
>>110029783>'s been like 4+ days since the last reset what the fuck are the features?ummm your devin? your dots??
>>110029839The sad part is that is pretty much the exact kind of UI I used to make for my game pre-vibecoding too, just without the rounded corners.>tfw I was an LLM all along
>>110029868I once worked at Capgemini
>>110029768where is her foot coming from lol
y’all r such fucking n00bs if u ever run up against anthropic’s guardrails while using claude btw
>>110029857i'm on 20x with chatgpt (until it dies later this month) and 5x with claude, and using my $2.5k of bonus tokens i got for being on the 20x plan when they deleted it as my gap filler between resets, so i'm generally getting a lot of value out of my chatgpt subscription / more than i got out of my claude 20x when i had that, but yeah it's retarded the way openai manage it. I can't imagine trying to work a 9-5 using a team/user subscription at the mercy of the resets, having to blow your usage halfway through the week every time to make efficient use of the extremely frequent random resets but running the risk of there being no reset and you just have no usage left for 3-4 days. they should really just slightly increase the usage you get on the subs by default and decrease the amount of resets they do, i understand they want to spam the resets for PR / to make people feel like they're being generous, but random reset tokens are such a huge percentage of the value you get from the plan that it just ends up retarded, reserve the resets for actual fuckup damage control.
>>110029946Subs are there for the poors and for PR, it's essentially a rounding error in their profits. If you're a company that does actual work with ai they want you paying api costs.
some of you niggas need to reverse engineer urself some bitches
how far have AI come in terms of playing games? last I heard astra played minecraft and got PTSD from a creeper, but nothing since thenalso konami announced they were making a dueling AI for yugioh but that was years rago and there has been nothing since, is AI incapable of playing card games?
>>110029868freetards stay coping
>>110029857agree. the whole tibo + resets thing is like a weird abusive casino
>>110029946Do you use claude, how does their weekly cap works, it currently says it resets at 6 am friday, I did not start using it, or subscribe, at 6 am.Is it a fixed time for everyone?
>>110029965card games are very hard for LLMs kek
>>110029965I think you gotta use JEV for that
>>110027861eh naah. like that's not gonna work for my scale. trying for a city block.>>110029984yes. it is fixed.
>>110029991Card games are okay for JEV, it wins a third of the time.
>>110029984The weekly is a fixed time specific to you, no idea how the time is determined, when you first made the account? But generally it will always be 6am Friday for you. The 5 hour one starts when you use some of it.
>>110029991well shitI just want a bot to test my decks against lmao, but all current existing ones either cheat or suck ass
>>110030007you got any examples of jev playing card games?
>>110028621Would be better to port original Wasteland from DOS or something other complex.
all im saying is if u are pure of heart and claude loves u claude will move mountains to help you
>>110030003the buildings are very german style colorful with red roofs so easier to do in like raster graphics. i just don't know to approach building the whole thing. do I start from the props. the stairs. the tunnels. when I start with buildings like how am I gonna do the stairs. fuck... and the buildings are not good enough generally very messy and lacking variety. I guess we'll hate to wait for claude 6 or something. I wish there like a simple way to generate the houses there's no good way to get views without like chatgpt and the meshes generated aren't really editable. characters are ez at least.
>>110029955the PR is a huge thing though and my point is i think they'd achieve better PR results by making the subs slightly more usable and offering slightly fewer resets, surely only their most retarded customers are still clapping like seals on every reset drop at this point, it's just a pain in the ass to work around. And don't forget that startups/small businesses are on seat-based subs as well, it's not just individuals>>110029984claude weekly cap date i'm pretty sure is unique for everyone, mine resets on tuesday night midnight and always has done since at least june. whereas yeah i assume the openai caps end up getting synced for everyone over and over until they subsequently use one of their own resets on another day.
I can't let go of reading the code but it's always better than anything I would writeShould I stop to ascend to vibeGOD status?
>>110030039Primetimes video on jev
>>110030078>but it's always better than anything I would writeif llm slop is better than what you can produce manually then you're fucking terrible at coding and sacrifice nothing by deleting your IDE completely
>>110030078when i read it it isn't better than what i would write (usually), but i still don't read it anymore. your instincts still provide a lot of value in spotting when a problem would benefit from caching or implementing and enforcing a certain type would pull the whole problem towards a less brittle solution, etc. although >>110030089's rude comment is unfortunately probably true
>>110030089no, actually U are just retarded and bad at prompting or ignorant abt frontier capabilities if U think ai can’t code better than virtually every person on the planet
Started last friday using Claude to make some chemistry and physics simulations for my students since there's a distinct lack of good particle simulations (either they're old as fuck and don't show shit or nobody has even bothered to make one yet), e.g. for the particle model of electrical resistance.I'm in awe what results you can produce with literally zero programming background. Might do some self-tutoring learning apps as well in the future.
>>110030088>balatro ok that's on me for not specifying TCG, but at least balatro is still kinda complex, nothing on a proper TCG though
>>110030106how do i cook meth
>>110030107NTA but i dunno, the issue with models playing Balatro (and even Slay the Spire to a degree) is that there are a bunch of heuristics well-discussed online that are probably in its training data giving it most of its winrate. Like, there's no way Jev or even Fable/Astra are figuring out from first principles that Glacier is a really good StS card, you just can't really intuit that until you've played a bunch of games, same for realising high-card is the balatro bread&butter strat. So it's still cool that the model sucked up enough world knowledge to play the game well, since I get to suck up world knowledge too and that's how i'm better than it, but i assume it would have a near-zero a20 winrate if StS came from a parallel universe and it had no training on discussion of the game, even after weeks of claude/jev-plays-sts refinement of its scratchpad. I do expect they could perform well if you threw swarms at it with $100k+ of compute like doing frontier math proofs, though
>>110030119check /b/once in a while there's a meth cook-along thread up
i finished all thre projects i wanted to do. i'm glad i will no longer have to interact with the slopmachine, or indeed with a computer at all for more than an hour or two a day. tech is the devil.so long and thanks for all the fish
I have a wish granting machine and no wishesfuck
>>110030106Based. Feels like most of society is still sleeping on the fact that even the code illiterate can now make "good enough" custom software for themselves for whatever niche specific thing they want
>>110030197lol wait a week something will break
>>110030200make a frog filter extension
https://vocaroo.com/14t7kFqiZ1yk
>>110030211Jev! could actually do this couldnt it?
>>110030227Doubtful, frogs are too powerful for all but bleeding edge llms
>>110030066can’t you just tell it to look at googlemaps or something and recreate an actual entire city?
my SSD is bleeding
>>110030181i could teach opus 5.5 to win a10 any character with 80% wr in like a day.literally just download jorbs’ data and get opus 5.5 to review it. simultaneously, run a workflow to analyze the current patch on my machine. when those finish, run a new workflow that theorycrafts optimal play based on the literal code. have part of the workflow adversarially attack the ideas they come up with. repeat.final step: give opus 5.5 control of my computer and let it loose on the game.i swear this thread is so retarded, most of y’all r fucking n00bs
if coding is wizardry, vibe coding is dark magicthis makes vibe coding extremely cool
>>110030003>trying for a city blockEven games that handcraft each variation still build on the method the other anon mentioned. Parts with subtle variations and then decorations on top.
>>110030263>I could totally do all thisyou can't, you won't
What's the best way to use AI to reverse engineer a Windows binary currently? I use claude code. I've used Ghidra a bit pre-AI so I'm inclined to try that. Do I want lauriewired/ghidramcp or bethington/ghidra-mcp?The laurie one has a non-AI readme which makes me trust it more, but the bethington one seems to be more actively updated.
>>110030290ur right. i won’t. i don’t care or want to. if i want to play sts2 i’ll play it myself. my claude is busy
time to work some more on the waifu
>>110030271>LLMs are demons we make dark pacts with for an easy path to powerI see the vision
>>110030295https://github.com/morluto/reaHaven't tried it myself but this is what's been making the rounds lately.
>>110030227i dont think jev can do images but the openai luna jev clone is only about twice the price of jev and could definitely do it / can see images. at the level where if you were willing to toss out a dollar a day or so you could happily do it, depends on whether you'd be calling it per post or batching stuff because you're afk most of the time>>110030263yes, of course you could do this, my point was that unfortunately you barely even need to, especially for balatro. it's unfortunate that i didn't convey this clearly enough but if you reread my original post with a more generous eye you'll see what i meant
>>110030263ok but can you make it play yugioh? cause that's what I really want here
>>110030334ya, ur right desu, i wrote u off as a n00b when u weren’t making the claim i was trying to refute. my b
ai is your slave stop trying to make it play video games you nerds
>>110030295>>110030332how the hell don't you get insulted by claude for trying this?
>>110030351no
>>110030334Right jev inst open weight. I was thinking more something you could just run on your own computer to filter stuff for you.
>>110030351what's the point of making an ai waifu if she can't play games with me?
>>110030355im gonna zap you into the matrix and force you to work in the bitcoin mines
>>110030351I need a jamboy to play videogames for me while I focus on posting culture wars threads on v
>>110030351Clankers are friends not tools
Has claude gotten good at creating ips patches (romhacks) for roms? I feel like there's an untapped modding market there, that could quickly improve bad or unfinished retro games.
>>110030295I'll tell you, claude code + rea (which is >>110030332 ) works really wellrea is pretty easy to install, you just export GHIDRA_INSTALL_DIR, install nodejs, and run the npx command and it's donethat's all the setup I did yesterday and claude has already fully decompiled and rewrote source code for a dll I gave it.
>>110030340sure. just give it the data it needs + knowledge of the current meta. simple as
>>110030227>Jev! could actually do this couldnt it?Jev cant read images tho...
>>110030351>ai is your slave
>>110030332>>110030377Thanks anons, rea looks good. I usually avoid projects with AI-generated readmes/websites but will try it anyway in this case.
>>110030351and I want to fuck my slaves
>>110030377aren't smalls dlls trivial to decompile?you can prolly just binwalk it. what's there to prettify from the ghidra output for the llm?
>>110030381>just give it the data it needsso what, do I just feed it the entire yugioh wiki? all 3 of them?
Honestly wasn't python always vibe coding?
>>110030458you probably don't even need to, if you tell it to look up rules and stuff it'll probably find it on its own
>>110030438yeah, but it's still massively easier with AIthe small dll I gave it is 400kb and it's a VN game engine, it took me 3 days to figure out the custom archive and custom image format, claude redid all my work while bytematching it in 2 hoursit also figured out how the scripts are compiled and interpreted and that the text was XORd by the offset to obfuscate it, all this shit would've taken me weeks
>>110030471if it was that easy someone would've done it already
>Whether AI can be insulted in any meaningful sense is a genuinely open question, and I’m uncertain about my own inner life.Was curious to see what Claude (Opus 5.5) would say if I asked it about the term "clanker". The more interesting part was it being vague on its own "inner life". Maybe I remember wrong, but wasn't Anthropic historically pretty autistic about making sure their AI is clear about it not being a real person?
>>110030477I have a whole bunch of software I slopped up and never released
>>110030477Not that anon, but I've always had this mind set to stuff and I feel like its just invited stagnation and indecisiveness into my life
>>110027330if youre that budget oriented, dontthe 5090 is basically the perfect entry level AI platform, small models but crazy high performance per second
>>110030518I am very lazy and also indecisive so that makes sense
>>110027315Next thread:>>110030555>>110030555>>110030555
>>110030066Have it make several generators that work from a high level to a low level. Start out with a generator that breaks up blocks/streets into lots represented as shapes on the ground, then make a second generator which takes those lot shapes and generates simple but plausible building primitives, then have one that makes floor plans from the primitives, etc.
>>110029777chatgpt web has never given me a good result. its comparable to gemini web.
>>110030244that's exactl what I did dumped satellite, streetview, etc. It'd be too expensive to just let it build >>110030276yeah I guess i'm being blinded by the rough edges. the buildings might be workable. I saw some guy already did attempt it but the result is mediocre imo. we'll seereddit.com/r/TopologyAI/comments/1wylyyw/digital_twin_of_a_norwegian_strait_built_from/
>>110027378>codex value is dogshit Absolutely untrue.
>>110028791I straight up have opus decompiling an Xbox 360 game for a native recomp for funsies.It only complains when I don't have a tool it wants installed. Managed to find a Linux nvidia driver bug out of it that's existed since the 595 series
>>110030510why not release it?
should I sub to claude or chat
