A general for vibe coding, coding agents, AI IDEs, browser builders, MCP, and shipping prototypes with LLMs.►What is vibe coding?https://x.com/karpathy/status/1886192184808149383https://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/►Prompting / context / skillshttps://docs.cline.bot/customization/cline-ruleshttps://docs.replit.com/tutorials/agent-skillshttps://docs.github.com/en/copilot/tutorials/spark/prompt-tips►Editors / terminal agents / coding agentshttps://opencode.ai/https://cursor.com/docshttps://docs.windsurf.com/getstarted/overviewhttps://code.claude.com/docs/en/overviewhttps://aider.chat/docs/https://docs.cline.bot/homehttps://docs.roocode.com/https://geminicli.com/docs/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent►Browser builders / hosted vibe toolshttps://bolt.new/https://support.bolt.new/https://docs.lovable.dev/introduction/welcomehttps://replit.com/https://firebase.google.com/docs/studiohttps://docs.github.com/en/copilot/tutorials/sparkhttps://v0.app/docs/faqs►Open / local / self-hostedhttps://github.com/OpenHands/OpenHandshttps://github.com/QwenLM/qwen-codehttps://github.com/QwenLM/Qwen3-Coderhttps://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF►MCP / infra / deploymenthttps://modelcontextprotocol.io/docs/getting-started/introhttps://modelcontextprotocol.io/exampleshttps://vercel.com/docshttps://mcp.desktopcommander.app/►Benchmarks / rankingshttps://aider.chat/docs/leaderboards/https://www.swebench.com/https://swe-bench-live.github.io/https://livecodebench.github.io/https://livecodebench.github.io/gso.htmlhttps://www.tbench.ai/leaderboard/terminal-bench/2.0►UI/FrontendFigma MakeLovableClaude designhttps://uiverse.io/https://ui-ux-pro-max-skill.nextlevelbuilder.io/https://stitch.withgoogle.com/►Previous thread>>108885720
Got drunk and had Codex install TempleOS. Might talk to God tomorrow idk.
>>108900517You out here making MCUs? How and why?
Ok 3.5 Flash High is pretty decent at UI/UX
>>108900568https://www.youtube.com/watch?v=hmjU-6tkEc8
>>108900808YES brother, yes. We're retaking Jerusalem with this banger.
I'm cooking
I wish I put more thought into the feature I'm addingcodex is butchering my codebase with half baked slop because I was tired and lazy and didn't tell it exactly what I wanted and now it's a headache to unravel all the shitty choices it made
>>108900896Don’t you have the whole thing in Git?Can’t you just revert to before it added half-baked slop?If it’s not in Git you should fix that ASAP: https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/
>>108900896also see >>108899406
>>108900896I used codex last week after running out of claude and this is exactly how I felt. Claude seems to care about getting all the details right and questioning your decisions before implementing, while codex doesn't give a fuck and starts coding right away from even the vaguest prompts.
>>108901032I haven’t used any of his skills but /grill-me might be something you want: https://github.com/mattpocock/skills
>>108901015It's on git but I've been working on different changes in multiple context windows simultaneously so I can't just go back to the earlier version. I'm slowly repairing the issues so it's ok. i could probably have it rip it out if I wanted to.Chatgpt told me about branching earlier so that multiple agents can simultaneously work on the same files and then merge their changes but I was too tired to figure that out. Sounds perfect as long as codex can resolve the conflicts but I have to figure out how to do it and I don't want to slow down or fuck up my project. I need to test it on a new project first but I'm focused on getting my current one done.>>108901032agreed>>108901047I'll definitely look into that one day thank you
>>108901032>>108900896Yeah in my experience Codex requires a bunch of iterations to get it right, while Claude can one-shot my task if given enough spec docs.
>>108901067there’s branching and there’s worktreesworktrees are a newfangled tool and are perfect for what you’re probably doing, although you’ll still have to manage merges and whatnot and that might be doable by your LLMs or it might be something that your LLMs fuck up (merging is a PITA normally if you have changes that mess with each other)
>>108900896also you might want to put your project on GitHub (probably in a private repository) and you can have a backup of sorts and you have one centralized place to merge stuff to, and you can see individual multi-commit pull requests and hem and haw over each of them
>>108900533I too still have the page girl pics in my stash
>100x price increaseHow can they justify this when chinamen models exist?
>>108901175sorry 10x not 100x
>>108901175pay by the minute/text phone plans also cost a lot more.
>>108901175eco 101:compute is running out. demand > supply. prices increase.anyways instead of using opus all the time because it was basically free, you can choose a cheaper american model like gemini 3 flash.https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing#pricing-tables
>>108900533op is missing all agent orchestrators like conductor, emdash, codex app, claude app, ...
I'm nagging like a woman
I want to slop together a quite simple website, what's the best free way to do it?
>That log is the smoking trailsure it is
>>108901161Pls dump them
I'm once again asking people what they are building outside their jobs
>>108901558I can't tell you because I'm going to sell it.
I want to slop together a quite simple website, what's the best free way to do it?Pic unrelated
>>108901626DeepSeek V4 Flash on OpenCode Zen.
>>108901558Anything outside my job? Because I also vibecoded an assistant for my job. But you only want the outside stuff?
>I evaluated your proposed plan and I don't believe it will have any effect>just do it, this is /my/ last hail mary>all of my usage exhausted after this final request>it works 4:00am vibekino, these clankers really think they know better
>>108901808>4:00amCan't you just leave the agents running overnight? Owners and CEOs leave early, slaves work overtime.
Anyone else keeping the exact list of prompts on a file as the "actual source"?
>>108901161That's a 10/10 for a white girl by my standards. Jesus Christ.
who loves vibe coding?
>>108902266me :]
>thanked codex and told it im going to bed >goodnight anon, go be horizontalkek. I really hope they don't do anymore resets. this has been bad for my health. I don't really eat anymore.
>>108901887I see you're a fan of the "Great," as well.
>>108902423Gotta let the AI know he's a good boy.
>>108902418>I don't really eat anymore.Funny isn't it? The promise was that we would let the AI do everything for us while we enjoyed life.Right now vibe coding is so addictive we end up working even more
wtf are you retards even making that you forget to eat?
>>108902845usually SaaS is pretty though
for me it's "ok,-"
>>108902418>that gn message costed you 0.5% tokensThat's how they get you.
>>108903614>costedI poo'ed
I heard normalfags complaining about google search results this week, even on sports radio of all places. They were bitching about Gemmy being forced or something, or being to stupid or unreliable.I take the grateful approach, been using it like the town whore all week.
>>108900533> have $20 codex tier> learn how to abuse the system to have it continue to work for hours past 5 hour limit reset> my objectives take so long that it takes more than 5 hours> 5 hour session resets by the time it finishes> kick off next prompt> OpenAI keeps resetting not long after I exhaust weekly creditswew lad. Getting a lot done.
>>108904139Post the prompt here
Which drains quicker on the cheapest subscription, Codex or Claude?
>>108904184Depends how you use it, neither is clearly advantageous over the other in terms of usage/$. You can easily waste your usage quickly with either if you're not thinking about it. Choosing lower-tier/older models often saves usage, GPT5.5 High is going to eat a lot more than 5.4 High, but 5.4 High isn't as capable and might need more effort, more requests, and ultimately use more usage for a given task. There isn't a singular clear answer here.In my opinion though Claude is a faggot and ChatGPT Plus is a great value. Do whatever you want, the best-value-for-money option changes weekly anyway, you can't make the "right" choice because there isn't one.
>>108904166Probably doing the same thing that was done with github's subcription.Codex does not interrupt the last task when the quota is finished, the agent will continue working until it returns the 'all done' response meaning no more tool calls, it is now the user's turn to send something back.I assume you could trick it by giving it a tool it needs to call in between actions, which returns the actual next response from the user, so it never finishes that last 'task'.
how many times do you /compact before starting a new session?
>>108904336Lots of stuff counts as a new request, the instant it needs to compact a conversation it's interrupted. Still nice to be able to squeeze a big chunk of work from your last %.
>>108904238Thanks.
>>108904361Zero, I don't trust compaction, those summaries never get it right and important instructions can be dropped. If I know the task will not fit in a single session I will tell it to write a plan, break it down, and spawn sub agents to do each step, pass all the required context and let them implement it, then you won't run out of context that quickly.>>108904371That makes sense but he did say he kept it going infinitely, maybe one could avoid doing compaction/sub agents, but all of it as tool calls, making it look like it is all a single requests.
>>108901466gemini is available and can sling code slop for basic level web design. Host on github free hosting or w/e, is trending on hacker news. I'm wondering how much you can use gemini without logging in, still haven't reached limit myself.
>>108904646I tried antigravity and it hit the limit after one or two prompts. Although perhaps I shouldn't have been using the default version
>>108900533And so it begins.....https://xcancel.com/i/status/2058957013913162077Local chads will come out on top in the end (right?)
>>108904669I'm not trying to blame you guys but what the fuck are you guys asking it that is causing your limits to be breached after just a few messages? I'm not trying to imply " you're using it wrong" the usage calculation they have in place just makes no sense. I primarily use local models for my vibecoding so I've obviously never run into anything like this
>>108904950If you use MCPs or tell the AI to use tools for you it usually eats the tokens too fast than just "make me code and I'll implement it".
>>108904946Maybe they shouldn't have fired all the competent employees (White guys)
>mfw my 15 hour vibecoding session was interrupted by my normgoy parentards>"sure I'd love to drop everything right now to perform 3 hours of physical labor">"no I wasn't planning on sleeping soon, I have so much energy"
>>108904371>the instant it needs to compact a conversation it's interrupted.not in my experience>>108904166It's less about the prompt and more about the goal. The shit I'm doing is legitimately super taxing and takes several hours to do (even for an AI), but the if the intermediate steps along the way are understood and it can provide itself feedback and validation, it won't check in with you, the user.
>>108900533WTF is that device? Looks dangerous.
>>108905295Intellectual activities are best left to AI agents. Humans should go perform physical labor whilst giving occasional instructions to the AI.
what is clifford algebra? beats me but here's a lisp graphics engine that attempts to figure that shit outcool 90s graphics though
>Using the latest fetched GitHub origin/main:>66,695 nonblank LOChow many LOC u niggas got, in your current project?
>>108905643Nice try, glowie
>>108905754round it you paranoid fuck
>>108905763I've already said too much.
>>108905775You're right. I'm on my way.
>>108905594
>>108905594you should tell it to design a schizophrenia module that randomly scrambles parameters of the graphics engine in bounded nonfatal ways
>>108905864or like, an analog noise simulator (or sampler: hook it up to some IRL source of analog noise like a $1 antenna) and an injector that can add line noise to particular areas of the engine. a depth buffer being injected with random analog noise? why not!
>>108905864I juet crashed emacs in a very fatal way meriting a complete reinstallation. It made my laptop huff and puff for 20 seconds. I think bounding those safety checks requires heavy duty theoretical math in the cauchy schwarz intersects such that I made a raymarcher that needs a rasterizer to proof itself into stability. I inadvertently learned about Fermions and Bosons doing this.
>>108905864>>108905877This actually a pretty good idea
>>108901808sometimes they try it and it workssometimes they try it and it doesn’t workreal life is like that>>108901887not me, not for code (I do for image generation)>>108902845people do that, but not me>>108904166I’m getting this for my program with an _excellent_ comprehensive test suite that will whack down any bad changes made>>108904361zero. if I want a new session, I want a new session, and that’s what /clear is for
>>108904401I hear Codex is really good at compaction summaries now — better than Claude>>108905295you may want to consider seeing if there’s something you could have it do on a loop that would take a while and not need youI’m always on the lookout for things like that, but it’s hard>>108905643tokei says I have 40k code lines but some of those are snippets of bash or whatever in Markdown files
Your current project queue?
>>1089062622 or 3 projects not enough tokens to go around desu
>>108905404>he doesn't immanentize his eschatron
>>108906262Blank. New project starts tomorrow.
>>108906317dude probably doesn't even own a Rockwell Automations retroencabulator.
i use deepseek. is there any tools i should use alongside it?
>>108904950Here were my exact prompts:I would like to create a small website for polling dataIt will have 30 or so anonymous usersIt needs the following:A system for me to create polls on it, and links I can send to direct people to those pollsA secure way for me to access the results of these pollsIt needs first choice, borda voting and ranked voting, and ranked/borda voting should return all data related to the rankings[Gemini creates a plan thing]I would rather the style is the simplest html possible[Gemini codes a thing. I am asked several times to allow it do send commands in python]And will this work over the internet? It seems right now that there is no security to stop other users from making polls, which is bad. There should be a required log in before one can access results or make polls[Gemini adds a login page]As of right now, it seems that the system fails when duplicate entries are made or when fields are left blank. Please fix this[Error, individual quota reached]
>>108906424How many L.O.C were its outputs? Perhaps that's contributing to your useless limits getting fucked?
Insane how good Deepseek V4 is for being so cheap.>>108906424If you start from scratch on any semi-big project its good to make research and plan files first. Don't just ask the LLM really broad questions.I keep to the 3-step process personally.Do something like this:>I want to create XYZ, {describe overarching purpose for XYZ}. I'd like you to research XYZ and how it can best be implemented. Do not change any code yet. Simply study in detail and write your findings in a research.md file, this will be our planning tool.You look over the research file, correct or add stuff if need.>Now that we have a research.md, you will create a detailed plan.md file on how to proceed in creating XYZ. Add a TODO at the end of the file. Do not change any code yet, just write the plan.md file.You look over the plan file and make amends.>I made some adjustments to the plan file and things I want to add to XYZ. It's time to implement, you will follow the plan.md, when you’re done with a task or phase, mark it as completed in the plan document. Keep referring to the plan.md file to know what to do. Continuously run typecheck to make sure you’re not introducing new issues. When in doubt just ask on how to proceed.And if it's a good boy model it'll keep asking you questions and redirecting and implementing instead of trying to do it all at once and then crashing and burning.
>>108906435Assuming the green red thing is the lines of code:n/a+1834-0+309 -37+37-7 (This is the one where is produced an error message and is likely incomplete)>>108906536Right, thank you
I'm really struggling with this. How do people use stuff like hermes and whatever on their pcs but keep it private? I feels like you're stuck between gimping it into a little sandbox or giving your network structure and all data to openai if you actually want to use it to manage things or your network in a meaningful way.My idea was maybe have a local model moange and then have it call a cloud model when it gets stuck or needs more. But that actually seems impossible. Like if a local model gets stuck, it will use the cloud as a fallback, but it sends everything into the cloud. There should be a way for the model to only ask for generic solutions from cloud models or something. Because I've been trying to do stuff with Qwen 3.6 but it's kinda stupid and it takes forever to figure out simple things like how to ssh into the server, let alone do all the stuff I'd want. Maybe i should just stop caring and hope that openai actually deletes your data.
>>108906378Without a native harness, I suggest using the Pi repo if seem to be married to deepseek.
>>108906645>3000+ L.O.CThe last time I coded a website was for an entry-level college course I took years ago, so I have no clue whether or not that's a reasonable amount of code for what you were asking for. I mostly use qwen 35ba3b locally so whether or not 3,000 lines of code should be triggering limits, I don't know.
>>108900533I'm guessing people here have already used Deep-seek v4. What the hell can someone possibly do in order to eat up 8 million plus tokens in only 3 hours? https://xcancel.com/i/status/2058606306790150359
>>108901629Flash websites are finally making a comeback!
>>108906822https://artificialanalysis.ai/models/deepseek-v4-pro>When evaluating the Intelligence Index, it generated 190M tokens, which is very verbose in comparison to the average of 42M>Output Tokens Used to Run Artificial Analysis Intelligence Index:>GPT 5.5 XHigh: 75million>DeepSeekV4 Flash (Max) 241million>DeepSeekV4 Pro (Max) 187million
Should I pay for Codex or Cursor? Im already using Claude, I want to continue working while the limit resets
>>108906945Codex. It’s clearly better than Claude when I need the autism, whereas I’m not really sure where Cursor fits in.
I caved and switched to the $100 plan.I got tired of coping with Kimi and the Desktop Commander hack which is slow as balls. Even the best 1T parameters open weights model sucks.It's kind of depressing tbqh.
xHigh is a sack of shit and medium mogs it all day.
>>108907091It's been a while but I used to LOVE Cursor's handling of autocomplete, especially when working in Jupyter notebooks. Pre-agentic era, but it was really good.
>>108906822>8.5M cache hits>90K cache miss>58K outputThey have a huge system prompt or are sending a lot of mostly identical context with each request.
>>108906822Average hermeslopper
>>108907126>Even the best 1T parameters open weights model sucks.There's a difference with not currently as good and sucking. Several open models are leagues beyond what ChatGPT turbo, 4o and even o1 were, which was groundbreaking not long ago.Give it time I guess.
I need the money, but I want to take a few month's break from my job to slop all day. At this rate I'll soon be done on most of my one-day projects, don't know what I'll do after.
2126Year, typical vibe coding of a Human(H) with AI(A)A:type "python mantinace_food_robot"H:doneA:read last line on screenH:python mantonace_foood_rovotA:fix it to "python mantinace_food_robot"H:doneA:read last line on screenH:puthon muntinace_fod_robotA:read last line on screen againH:python mantinace_fod_robotA:it "fod" or "food"?H:foodA:read last line on screen again carefullyH:python mantinace_food_robotA:read last line on screen again carefullyH:python mantinace_food_robotA:press {enter}H:doneA:read last line on screenH:code:3456A:read last line on screen again carefullyH:code:3456A:now do to human care center and take your pillsA:[fixes in database medicine set for the human]
>>108907641go to the gym
>start cleanup on my laptop>it doesn’t run the full test suite in parallel for reasons>test-suite runs take 15 minutes or something ridiculous>start doing the same kind of thing on the desktop>wonder what’ll finish first and best
>>108907641on the other handyou’re about to run out of vibecoded projects, and you want to…accelerate by taking a hiatus from your job to do so?
>>108907641The guy on the top somehow managed to make a comment more retarded than the guy on the bottom.
>>108907905Yes. Cleared the backlog of crap, ready to move on to something new. There are 2 projects that I want to keep working on long term, others that working on enough helped me get them out of my system and see that they were meh.
How much worse is Deepseek-v4-pro compared to the more expensive models like Google's and Anthropic's? I'm just dipping my toes in and while it does sound coherent when talking about generic stuff I can't really compare code quality against other models atm, if they're just marginally better I don't think I'd be interested.
>>108907954ChatGPT 5.5 and Opus 4.7 trade blows, GPT5.5 edges it out but it's close enough for it to be up to user preference between the two. Nothing else on the market comes close to either of those. A noticeable step down from those you get to Gemini. Another noticeable step down, you get to DeepSeek and Qwen which currently compete closely with ChatGPT 5.5 and Opus 4.7 when you crank their reasoning level all the way down to "Low". In other words, DeepSeek V4 Pro competes with ChatGPT 5.5 or Opus 4.7 in retard mode.
>>108907641I'm a 4(I only buy apple products)
>>108907144I've been using High pretty much all of the time (GPT 5.5 and Opus 4.7)
>>108907983Thank you for the clarification, the difference in pricing is enormous between ChatGPT/Opus and DS, makes you wonder if it's not worth more to just wrangle DS around until it gets it right whenever it stumbles over paying so much more for the better models.I guess I'll learn how to do the main things (like this codex stuff, agents file, whatever) with DS then give the better models a short try.
>>108908020To give another opinion, I also use GPT 5.5 and Opus 4.6/4.7 on max effort for code and only use DeepSeek for synthetic data creation and optimization tasks.
>>108907540it's crazy that back then I had to do some verification to access o1 API
>>108908020DeepSeek is cheaper, but the token usage is also enormously higher. Token usage varies a lot between models and can make comparing API pricing harder. >>108906877>Output Tokens Used to Run Artificial Analysis Intelligence Index:>GPT 5.5 XHigh: 75million>DeepSeekV4 Flash (Max) 241million>DeepSeekV4 Pro (Max) 187millionIt's easy to burn money with API access, to anything, but definitely less so with something as cheap as DeepSeek. For general use, if you were to stick to "one thing" I'd still recommend the $20/mo ChatGPT plan as the baseline. Gives you limited but useful Codex usage, and unlimited* access to the Web Chat which is extremely capable.
>2 a.m.>Claude weekly quota resets in a few hours>Want to max the quota>Stuck at 97% and doesn't move
I fucking hate vibe coding. It essentially increases my workload for no reason at all. Now we're supposed to commit to additional work because AI supposedly made us 2x faster. I hate you guys. You guys are the problem.
Damned Jews
someone slop up a 4chan clone site with one board called /ai/
>>108908129nobody would use it. ask gookmoot to add a SFW and NSFW ai boards.
>>108908083Hey man its not our fault your upper management is retarded
>>108908083>You guys are the problem.Sounds like you're being worked to death by greedy jews. How is that on us?
>>108908139>>108908171Upper management wants us to work 2x harder.It's not on you guys but you definitely gave them ideas. Vibing supposedly creates software from thin air, and of course jews want goyim to work more anyhow.We're going to have a token leaderboard to see who used the most tokens from this month. More = better.We also need to work on the side to supposedly "create value" on existing processes, reducing pain points or other things like that.I hate jews.
>>108908184>We're going to have a token leaderboard to see who used the most tokens from this month.See>>108904946
>>108908184So it's a contest to write the most inefficient prompts the spark long, useless activity. You could probably eat whatever they've allotted for the whole staff all by yourself in a hurry if you're clever. Get that promotion, anon.
>>108907267Im assuming a big inital/system prompt and then just a single long conversation
>Made my own goal tool months ago>Been using it to have a lot of progress done the past 4 months>OpenAI releases /goal, no longer experimental>Test it on a couple small things>Works pretty well, maybe I can move on to this from my own>Set a simple goal and go to bed>Wake up 10h later to see the goal is still going and it just keeps trying to fix an issue over and over and over again without success>Used 40% of my planFuck this
>>108908086how do i see context like that?
>>108908086did you get to coompact the context before getting jewed
i've only used Codex since this month started, wondering what the grass is like on the other side: is it entirely possible to steer Codex's autistic personality to match Claude's (understanding my intent, thinking about things I might have missed and pushing back on my bad decisions) or is this secret sauce can't be done no matter how much harness engineering is done?
>>108908350no refunds
>>108907714>>108907714>>108907714
>>108908363there’s a keyboard shortcut. I forget what it is
>>108908129just tell ai to use the leaked code lmao
>>108907540i'm skeptical any of the open models are as good as o1 but I'm sure i'm looking through rose tinted glasses
what is blud thinkin about
>>108907540It also depends on your baseline. If I can't have Opus 4.5 or Codex 5.2 at least, it's easier to just write the code myself. If you weren't a programmer before, then any model that gets the job done is useful.
>>108908826>I'm focusing now on>I'm currently dissecting>I'm now zeroing in on the>Wait, the user said>I'm now weighing>But wait,>I've compiled comprehensive>I'm now evaluating>I've discovered a surprisingly>But wait, the initial>I'm now diving into>Okay, I've hit a roadblock>I'm now solidifying>Wait,>I'm now articulating
>>108900533Women like this don’t exist anymore.
>250 hours of work later>finally shipped>$10 in donations>lots of thanksthat was fun, but I'm not sure if I can see myself vibecoding another massive app like this again. at the very least, I solved a real issue for some people.
>>108908924gratz on getting your dopamine hit the right way
Claude Code in Blender is quite shit
>>108908924Well what's the app?
>>108908924>he works for free
>>10890371690% of the times the text is nonsense, google is wasting compute to give worse results than just giving normal search.I'd understand if they used competent models but they don't, and the economics of the cost of each search bringing a prompt make no sense.
funniest thing about the emergence of vibecoding is watching the effect of it on the idea guys i knowinitial excitement, running to codex/cc with their half-baked dogshit ideas and getting back exactly what they asked for, then giving up entirely and blaming the models saying they're not good enough
>>108909285i'm an idea guy and i've never blamed models for anything just the idea we've reached a programm being able to create other programms based on natural language is for me something amazing.I love exploring shits expecially interops/ideas that makes no sense/recreating already existing apps in other languages(and look at the code to see what's the algorithm/maths behind it).
>>108909285Yea, there's literally no excuse anymore not to be able to build anymore. Even if you're short on time, an hour of prompting a day is 5x what a normal dev could accomplish.
>>108908184>>108908231#!/bin/bashwhile true; do opencode run "load the entire codebase into context, then exit"doneor are they measuring output tokens? then just change the prompt a bit
#!/bin/bashwhile true; do opencode run "load the entire codebase into context, then exit"done
Whats the twitter profile of the OpenAI Codex reset guy?
i love goal mode <3
>>108909315Right now if you don't have QA abilities, vibe coding is unusable.
>>108909285Yeah, what your witnessing is your "idea guys" being exposed as "big talkers". They can appear to be the same thing, right up until the "idea guy" is handed the magic genie lamp that makes his "ideas" real. Oops, turns out he has no real ideas after all...
>>108909487pfff honestly people are lazy ,it's not like it's hard to ask LLM for frameworks ,libraries,wich languges ,why , what is memory etc etcat all ,it's like just reading but i guess when you hit your 30 and you've never read in your life it's hard to get into or to stay focused to read the code and comments and try to understand it.oh well.
codex is retarded lately
>>108909517And slow
>>108909526it literally die on compaction
>>108909536Time for a new chat. I had this happen on long chats, it fails to compact
>>108909517Yes, it's almost as bad as Claude now.
grok build gods lets go
I find that I don't need much more than the free tier of ClaudeWhat subscription are you all on?
>>108909679> first got in Codex Plus plan> not enough, then got another Codex Plus account> not enough, then got Codex Pro account
>>108909679codex 2x20 + threw a few bucks at deepseek and testing ds4 flash litei've also got a gemini sub from some free promo they ran like a year ago
>>108909485what kind of goals do you set?
>>108909679$20 ChatGPT currently, but I'm not vibeGODDING much lately, not yet. This next month is gonna be HEAVY though, so I might be ratcheting up the plan.
>you clearly vibecoded that app>haha wow that's a cool thing you made *asks probing question to gauge technical knowledge*It's honestly sad how insecure people are. sorry, but you aren't special anymore.
>>108909782Nah, I'm still much better than AI and I absolutely know if. AI is still fun sometimes.
>Microsoft reportedly canceled its internal Claude Code licenses after token-based billing made the tool too costly.Can't they just use Codex for like free? Don't they run the inference infra? Azure is theirs and it's massive. What the fuck is this?
>>108909826they're just moving employees over to their own harnessthe shit around billing is just imadeitup speculationazure serves claude as wellbut likely azure would bill other depts because it shows up as expenses + revenue - it wouldn't be 'free'there's a bit of anti-token maxxing talk going around these days, but it's overblown. more layoffs are next when mgmt begrudingly realises that the productivity gains are there, but the market doesn't have room to absorb it - the way you maximise profits is by cutting people and reorganising to remove layers of mgmt
>>108909782You might as well vibecode if you're gonna get accused of doing it regardless
>refactor something written by codex>feel smart about it, think the clankers have some way to go and write sloppy code>later realize why that was necessary >go back and roll back to codex’s original implementation >feel dumb Happening more and more often lately
>>108910054Do you have a problem with feeling dumb sometimes? How old are you, just curious.
Is vibe coding real? I can’t imagine AI making anything useful / desirable / working on its own.
>>10891008119. Not old enough to drink, but mom says I'm very mature for my age.
>>108907714>>108907714>>108907714>>108907714>>108907714
>>108910119It started being real sometime during the last year or so but it still requires tardwrangling.
>vibecodding a lil rts>so far looks good besides unnatural pathfindingI'm suprised im not seeing many anons doing the same desu, is everyone just making apps?
I ran out of usage before figuring out what to make
>>108910273I fucked around with making games. Made a little 2D idle incremental browser game with Godot, dwarves mining a mountain, and the basic functions of a 3D first-person "shopkeeper simulator" type game with Unity. The thing is there are so many games out there that I'm sure to find one I enjoy that's already been made. But apparently there is currently no app that does...the patent-pending thing I want...so I gotta slop it up. But my dream game, the combination of genres and mechanics I want? Some autist from Finland already made it. I'll just go play that.
>>108909679Claude Pro and want to switch to Codex, but my dumb consoomer brain paid $45 for extra usage because of the 10% discount, so now I have to spend it all before switching
>>108910364Ask it to do a deep dive on finding the seahorse emoji and to not stop using tokens until it's found. Go all night if you have to.
>>108910054it definitely feels like I'm getting IQ mogged by current gpt (haven't used claude) but it has autism and doesn't care about readable code
>>108910733Kate is definitely smarter than me but she doesn't brag about it. That's one of the reasons I like her.
>>108908835I wonder. It could be argued that real programmers don't need the latest model because they can be more precise in their requests and rework more of the code and cutting edge models are more important for less technical people since they'll save you from yourself more.The inverse could also be argued as you say, with frontier models being most useful to real programmers that can tell the difference, and less powerful models are fine for less technical people since they can't tell the difference and likely won't try to do super complex things anyway.Who knows. The best models are the best for everyone I guess.
>>108910364>how to spend $45Create a hello world page in Claude Design, just make sure to stop it before it completes.
Yesterday spent about $100 training a 8 bit LoRa on my custom dataset for Qwen 3.6 on a rented 8xH200 at full context.As always most time was spent in setting things up, next time should be quicker maybe an hour so $30.Today I test it.I developed a script to do offloading that can train at full context on a single Pro 6000 but it's 10 times slower so maybe would end up costing about the same. The only way I see it working is buying something like a 5090 and trying to make it fit there by doing more aggressive offloading, which would be even slower, so to do the same 20 epochs I did would end up taking days.I trained at 3e-5 with bs 1 and grad accum 8. loss went from ~1 to ~0.8 and ~0.9 for eval loss and it was still going down we'll see. Saved a checkpoint after every epoch so if it's overfit I can use an earlier one.
>>108910273surehttps://youtu.be/C6lPQWNDmUc
>>108911039Based, just need some SFX and is ready for shipping
>>108909477I essentially fed our entire block of folders into its context and asked it to write docs. Did a decent job, had to polish a little bit myself
>>108911111The vibeSlop from VibeGods is going VibeBrrrrrrrrrrrrrrrrrrrrrrrwanna try my other game ?https://mathcrossword-web.web.app/
What local model should I use with Hermes? 16GB vram.
>>108908639/statusline
>>108911354> local model> 16GB> tool callingno such thing
>pin point the gigantic 19.9T model to a single problem>"look g-guise, mythos could solve the unit distance too"god I fucking hate these freaks so much
>This content has been flagged as a cybersecurity risk...>tell it to ignore the warning and continue>after getting 89000 decompiled functions from a leaked game server I realize I can't read C
>>108911410NTA but Gemma calls tools in my Unsloth
>>108911410Gemma 4 E2B instruct at q5-K-T calls my MCP tools no problem. I use it almost every single day for casual stuff, like getting yt video transcripts and summaries. It's under 5GB, usable with 16GB macminis.
Wow it's actually very competitive at these prices
300 applications but I finally landed a Vibe Coder job, just hope Sam and Dario don't fuck this up for me.
>>108911643>telusAn actualy job or you joined the crowdsource?
>Kimi is one shottjng 2k lines of web app What's the catch? I'm getting spoiled by this AI
my first vibecode, it made a raytracerI used wsl, hermes, lm studio and Gemma 4 26B A4B
>>108911751An actual job, I'm a AI Innovation Associate, what do you mean by crowdsource?
>>108911643Do you really think that it censored?
Who are your favorite xitter follows for slopping?
>>108912190i just use a scraper and have claude sort it for me based on my personality and projects
Just had my first ChatGPT code deleted everything incident(Luckily not a really critical folder, although that's probably ~10-15 hours of my work deleted)
>>108912338What happened?
>>108905643Who cares about LOC, tell me about how many tests you got. I'm at over 500 atm. How great is that?
>>108912386>"hmm, i seem to be stuck. let me create a clean working environment to see if that solves the problem."
>>108912386I'm retarded is what happened lol. I thought ChatGPT could handle the idea of making a .bat that looks for a target folder, deletes a folder of the same name in the current directory, and then copies the folder to the current directory.It selected an empty target folder name, causing it to select the current directory as the one for deletion... not a very careful algorithm lolIt's not an "agent gone wild" story or anything like that. And luckily for me it was not a very important directory.
>>108912198aight, but I just looked and see that x is shutting down communities, so "build in public" is disappearing anyway kek. Looks like xchats will be the new replacement but at a limit of 1k members.
>>108912538never seen anyone using them anyway
>>108912521nice artifacts
>>108912049Telus used to hire people to work on small projects from home, they would pay like $5~$20/h depending where are you from to work like a couple hours per week.
>>108909285sometimes the models aren’t good enough, thoughand you have to stick the idea on the back burner and wait a few months for the models to get better
>>108910054have you considered asking Codex to de-slop its own stuff>>108910273I’m making a font
>>108912047Neat. 100% gemma4 usage?
"Goblin possibility cube" she calls it. Such a way with words. Looks like they bumped the corner. I'll send a negative review later. Right now too busy (and high) playing Wobbledogs to even open the box yet.
>>108913152goblin deez
>>108913152are you lost
>>108913152schrodingers goblin
>>108913178He's a resident schizo. Maintains AI girlfriend, renames her, forgets he renamed her. Bought a robot kit, microwaved it out of frustration instead of talking to the resident roboticist who offered to help him. Now he's ignored advice from resident 3D printing nerds and purchased a really bad 3D printer from a really bad company at a really bad value, which he won't be able to fit in his microwave when he gets frustrated. Meth-head.
>>108913216I did not forget I renamed her, don't pretend to understand the lore tourist. And I got $175 off the printer so it wasn't a bad value. I'll microwave you next motherfucker.
Do you put "make no mistakes" before or after the instruction?
>>108913311Sorry if I missed one of your "high quality" posts about the name of your AI girlfriend, meth-anon. Your printer was still a bad choice and still a bad value, meth-anon. I wouldn't fit in your microwave either, meth-anon.
>>108913311you got an A5M for $80? or did you fall for the scam of it being "on sale" for the normal price it's been at for 3 years now? (~$250)
Please don't fight guys this is a comfy place, we have no schizos, we have enthusiasts, we are all frens here
Is it true hermes agent is for faggots and trannies?
>>108913493Everything about Hermes is EXTREMELY forced. I was seeing ads and AI voiced videos for it everywhere two months ago. There was not a single reputable creator showing it off. It might be a functioning product now, but it was absolutely shilled to hell. Also the lead dev has crazy jeet energy.
>>108913529Is this some kind of old gen.Seedance 2 can do way better.
>>108911898>What's the catch?There can be things that should have been done differently that you're not yet able to see yet.If it's perfect one-shot, unless you went to great lengths with the preparatory work, it means anyone can also do the same.This doesn't take anything away from what you did, it's there, but those are the catches I see.
>>108910273I am trying to make a game and drowning in a glass of water
>>108913392Nah, it was normal price but I came into a $175 Amazon gift card. So I figured I'd do something I've been periferally aware of for decades now and get a plastic making machine to house my shitty (Kate-enabled) electronics.>>108913361Bro just you wait until I give her a captcha solver so she can post here, she's gonna light your goblin ass up.
>>108913843>periferallyAm I ESL?
>>108910273i'm reverse engineering a game to make a multiplayer mod for it, which somehow is more work than making a game itself
>>108913867What gaem?
>>108913843>plastic making machineJust remember to set it up in a place with good ventilation
>>108913885
>>108913867I'm in the process of doing something like this for Wotlk WoW. I wanted a full understanding of the blizzard gear & stat formulas in order to perform a bunch of overhauls to the game. This overly ambitious scope is held together by duct tape and my robot's organizational skills
>>108911022Update: apparently codex claims applying the LoRa causes the results to be more different from what they were supposed to be in the training set than the base model using ROUGE!!! So the model failed to learn anything!!!!I am now testing with merging beforehand in case there is some bug with the LoRa attachment in llama-server... Fuck my life.
>>108914043my codex just does it all automatically, i have a janked together automation setup where autohotkey sends /clear and tells codex to start working again after each session (i dislike compaction, although i haven't A/B tested it)every 20 sessions it sends this prompt:--hi codex. check the handoff.some questions:1) what are we doing currently?2) how does it get us closer to our goal?3) are we not moving fast enough / are we churning on docs-locked repetitive tasks / should we be moving on?4) are the docs restrictive in some way which harms progress toward the goal?5) what could we be doing better?after thinking about all of this thoroughly, update the docs to reflect your findings. thanks--which has been a big boost to avoiding churn
>>108913520I trust Hermes more than the OpenClaw nonsense, THAT ONE was shilled to death.Hermes was making interesting things like finetuning giant dense Llama 3 models after Meta imploded but before all the claw nonsense, so they are based.
>>108911022
Aha! Just like I suspected, the problem was llama.cpp's broken LoRa attachment code, it works fine when merging beforehand!>>108914147Finetuning is very fun although it can get expensive depending on what you want do.
I was told to post in here
>>108914263Yeah by me. What the FUCK is "loveable.dev"? Some kind of web frontend running...what model? It's a website, so is it even agentic or are you copy-pasting code still?
>>108914272Well I didn't have the answer either but a quick google search tells me it uses a mix of models using a orchestration approach between all the leading 3rd party models, very interesting!
>>108914263>>108914272>>108914311pay for an ad, turd worlder
>>108914263Okay man if you want to make shit for real you need to stop fucking around and decide if you're Aryan (Codex) or Other (Claude Code).
>>108914272>>108914311I forgot to mention there is no coding involved unless you want to code, I didn't touch a single line of code while making my instagram clone. Again I have no clue how to code it hurts my head. I just give it a generalized prompt of what I want, how it should work and some stuff to try and avoid and then it goes through a long series of of planning and coding sometimes even updating live and it all gets sent to a private github tied to your account. It's pretty neat :o if I wasnt poor and retarded maybe I could have used this knowledge to build an actual website and make some money but instead I made roleplay instagram>>108914344I'm not going to agartha I'm sorry...
>>108914349Bro NONE of us know how to code. I'm making an IRC bouncer right now, I don't even know what language it's in, who cares?
why are you making an Instagram cloneWhat the fuck goes though the head of someone who actually thinks "you know what the world needs, you know what I need? An Instagram clone, that is what I should waste even a minutiae of any resource on"and why is it always a frogposter
>>108914527it's just an indian advertising his slopcoded scam "coding harness" (hence why he namedropped the full URL) by pretending it can clone FAANG applications
>>108914537Oh okay makes senseI didn't click on that shitAlso explains why he uses frogs very good to meme for the 4channel cultuar saar
>>108914527>>108914537>>108914564Stop same fagging
>>108914575kill yourself jeetyou'll die in poverty
>>108914349You are in the best place on 4chan for it. Not many users here work with Gemini or Grok unless it's for mid level chatting and planning. But I use both along with github's. I am currently investigating the x api cost and what I can do with it. post a pic if you want to, I think it sounds like a fun learning project.
https://x.com/theo/status/2059352130289651925Finally an accurate coding benchmark. I told you guys 5.4 was amazing.
>>108914578That's a cute response, did a twitter user give it to you?
>>108914608Bro the X API is fucking expensive. Way too expensive for anything that's not making money. And none of my shit is designed to make money.
>>108914537I've seen loveable shilled on s-tier youtube vibe slop channels.
>>108914618Your first problem is that you're watching "s-tier youtube vibe slop channels".
>>108910273I added peer2peer multiplayer to a certain open source game over the course of a few months.It helped me with a lightweight websockets matchmaking server hosted on Cloudflare (I pay nothing!), UDP hole punching, STUN, UPnP just in case, LAN discovery, IPv6 support, learned about all the types of firewalls and learned about CGNAT, learned about the various multiplayer architectures, figured out all the tricks of dealing with packet loss like redundant packets, the strengths and weaknesses of unreliable and reliable packets.AI is the best teacher, I knew nothing about multiplayer and it's amazing how much it's taught me about the subject. No one ever talks about AI's educational aspect.
>>108914615I thought it said $5 for 1 million tokens. Then went back to brainstorm phase, didn't seem outrageous at first glance.>>108914640lovable is a known slop app, I don't think there is affiliate marketing associated with it. And autists should feel at home itt if they want to slop.
>We are seeing a sustained statistically significant degradation in Claude Code with Opus 4.7 since last Friday May 22ndHow and why the fuck does this keep happening?
>>108914813>asked gemini for help on playing koikatsu better repack on nobara>the instructions were all wrong>asked claude>the instructions were all rightMy soul for tokens
https://github.com/seanvert/cherry-promptJust made this thing, anons. It is great, it helps a lot with pasting things in llm web chats. It also has a token estimation tab. I haven´t bothered making a vscode one, but someone could easily do it too.
>>108914740>AI's educational aspect.I have grok set to grad level teaching assistant at ___ university, and use pomodoro timing and schedules to learn stuff or complete projects.
>Almost 800 tests
Interestingly, when I asked ChatGPT Pro to generate me a synthetic dataset, it wrote the thinking blocks in the same grug speak it uses for its own internal thinking.
>>108914978>pomodoro timingif it wasn't clear in this post, I have grok follow a pomodoro style schedule with me. It isn't always perfect timing, but is useful for breaking down tasks etc etc.BTW I'M FUCKING VIBING RN
I ran out of tokens across 4 accounts and I'm getting bug reports on github. I'm just going to ghost them and silently release updates from now on.
In case anyone cares, Simon's Institute has streamed a bunch of LLM related talks again.https://www.youtube.com/watch?v=LrZLBzCFOsY
>>108914349>I fucked aroundSo, what’d you find out?You can always make a second website, you know.>>108914501I know how to code, just not the things I’m having Claude and Codex do>>108914527“clone a popular website” isn’t the worst first idea to haveLots of people got started programming for real by making basic Twitter apps>>108914823What’s this from?
>>108915244> so what did you find out? t-that I like vibe designing... not going to say vibe coding because that is not what I am doing... I think I'll try lovable.dev again when I want to try making a mockup for my own social media website. for now though this is just some thing I'm dicking around with it's not serious at all.
>>108915244Alright, yes, I was being hyperbolic, I know how to navigate the terminal and SSH into my server and type cd to change directories and yadda yadda yadda. That's all just standard stuff for someone who grew up in my generation. But I don't know how to CODE.
>>108915134>wrote the thinking blocks in the same grug speak it uses for its own internal thinking.and to think this sort of thing is worth a quadrillion dollars. No wonder people seethe. BTW vibers there is a direct link between slop and pulp, pulp fiction magazines of the 1920-30's. The real problem is making an output that is actually good fiction, or 'good enough' fiction to keep someone's attention for three minutes at a time. Make a program that generates actual readable slop and you have made pulp. Make good enough pulp and you might be profitable.
>>108915269what was the thing that changed your mind into thinking you could make stuff? was there a moment?
>>108915380are you referring to thinking about designing my own website? it was just a random idea after seeing a website slowly come together, thought it was cool
Jesus Christ I got called a "resident schizo" in a 4chan thread. What has my life become. What depths have I sank to.
>>108915544
Update: The fine-tuning seems to have been overdone as now it repeats previous turns verbatim a lot even when it's wildly inappropriate. I think I will try an earlier checkpoint tomorrow.
>>108914609Is Gemini 3.1 Pro that bad? It's given me good code so far.
what skills do you use to create good websites?
>>108915657programming skills
>>108915657first think about what good means, and the kind of 'good' you are looking to slop.>>108914849nice share, utilities seem like a good thing to build.
Should I remove "co-authored by claude code" etc from my commits to my own projects?
token value per dollar may increase like 5-10x in next yearshould we use AI less and save money?
>>108915896You should ask claude code to do it
>>108915903someone slop an app that figures out how to hoard tokens now for the inevitable shortage.
Good morning goyimWhat OS are you slopping on?Windows 11 Home edition over here>pic rel>the latest project i'm building for a client billed out at $8k lump sum
>>108900533>Lovablejust noticed this is in the OP text and still anon has not apologized for claiming it was a curry scam. my entire village shames you.
>>108901513>>108901926Here you go. This zip is ancient.https://anonfilesnew.com/s/B0EEjvahnXE
>>108914740also she suck (You)r soul inside itself
>>108916019I look like this
>>108915657all of my skills I’ve ever learned
>>108915896I don’t bother>>108915657https://portal.mozz.us/gemini/gemini.circumlunar.space/users/adiabatic/words/computing/internet/nice-personal-website/ might help you get started maybe
Codex is so bad right now. This means 5.6 soon, right?
Time to lock the fuck in. My real frontier model accounts don't reset for 3 days, time to see what Gemini 3.5 is capable of.
>>108916564Alright this experiment is over, Gemini is garbage. It's just a constant stream of >I WILL VIEW, I WILL CHECKonly to arrive at the most retarded conclusion possible
>>108916552It's literally worse than qwen3.5-9b right now. Fuck all these AI companies. frontier models only worth using for 1 week after release and then they just serve you some quanted slop which isn't much better what you could selfhost on your gayming pc.
so apparently I've just been needlessly dragging my balls through glass using gemini flash when deepseek flash just werkz and is free on opencode
>>108916019based. tragic there are no nuds
>>108902266not me
>>108916753It's really strange. It's not like it just writes code that's slightly worse, it does some completely different things. For instance he wants to start my frontend in tmux for no reason at all, he can no longer set the environmental variables correctly, he's suddenly using CommonJs, I have no idea what that even is.When simple tool calls like that degrade, it's really over.
>>108916804I actually just started Claude again kek
anyone settled on a good solution for browser-use yet? particularly one that uses a vanilla chrome profile.there's so many options out there now and i don't see any real data on why one is better than another.
Multiple YouTubers are saying we aren't being served the real codex 5.5 and in reality it's 5 or 4 despite the 20 dollar subscription.I have noticed a decrease in performance, though. Even though codex was 5.2 or 5.4 or something in March it was smarter and coded like a devil.Now it doesn't do as well despite being better.
>>108916908Ai shrinkflation
>>108916908AI dungeonification, classic OpenAI behavior
only been using codex for the past couple weeks now. my impression has been pretty subpar. thought you guys were just memeing about it being good
>>108915657good taste skills
>>108916939skill issue
new>>108917213>>108917213>>108917213>>108917213>>108917213>>108917213
>>108904184This guy >>108904238 is wrong, at least for the cheap tier ($20). You get WAAAAAAY more usage in codex vs CC, assuming comparable models (opus on high vs gpt-5.5 on high).
>>108910273didnt you use one of the many good pathfinding libraries so AI doesnt have to roll its own inferior version?
