A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://claude.com/product/claude-code (Fable 5 is the best LLM available, requires Max plan)https://developers.openai.com/codex/cli (Essentially scamming you with LLM degradation and resets that reduce your usage, but still the second best option and arguably the best bang for your buck in the 20$/month plan range)## Worth it for code, but the frontier models above are betterhttps://x.ai/cli## Not worth it for code, but maybe good for other thingshttps://antigravity.google/product/antigravity-cli----## Prompting / context / skillshttps://arps18.github.io/posts/claude-code-mastery/https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://github.com/mattpocock/skills — /grilling is a favoritehttps://github.com/DietrichGebert/ponytail## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/https://opencode.ai/https://cursor.com/docshttps://docs.windsurf.com/https://docs.cline.bot/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent## UI/Frontendhttps://www.figma.com/make/https://www.anthropic.com/news/claude-design-anthropic-labshttps://uiverse.io/https://ui-ux-pro-max-skill.nextlevelbuilder.io/https://stitch.withgoogle.com/## In-browser builders / hosted vibe toolshttps://bolt.new/https://replit.com/https://docs.github.com/en/copilot/tutorials/sparkhttps://v0.app/docs## Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.0## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109447707
morning saars
yann lecunn is sounding like gary marcus
>>109451817wtf happened to googlethey used to be the best until 3 pro
>>109450995I pay the 200 codex one for my own fun, I don't intend to ship or sell anything, it's just an expensive hobby and it's nice solving my local issues.
>>109451780>They openly stated as much, yes.Then why the flying FUCK to people still use them? Of all the shit you could feel brand loyalty to a model provider is probably the dumbest one to have. Literally just find a model and provider that isn't like this and use them. Why bitch and moan but then act like you HAVE to use them? Is Dario holding a gun to your head forcing you to get constantly cucked by them??
>>109451817Weird, I thought they were skipping that version because it's not really better than 3.1, or more like, it's better, but it's getting destroyed by every new sota model, open weights or closed weights.>>109451840In a vacuum it's probably an ok model, but not after the new opus, sol and kimi.I don't really think they'll sink or anything, they can pay their way out of this, just like a few years back when they were lagging behind everyone else.
>>109451817isn't this the guy who just does a bunch of random predictions and then deletes the ones that turned out to be false?
>>109451865They only sabotage you if you work on LLMs, or at least that's what they said.
I love Sol's desire to cheat on everything.>I found an excellent example on github, I'll pull it and incorporate it into our plan.>I've found a CC0 licensed example online so I'll use it instead of creating a new one from scratch.>The user has claude code installed, I could use that to get a second opinion.This bitch, I swear.
>>109451838I still appreciate his llm takes more than any dario or sam "let's fuck you over" word of the day.
>>109451889>The user has claude code installed, I could use that to get a second opinion.lmaooo
>>109451840I unironically think they are playing 4D chess. The moat will fall anyway with DeepSeek V4 Pro.Models are not the moat anymore. They have cheap inference and nothing prevents them from just hosting/refining chink models.https://newsletter.semianalysis.com/p/google-we-have-no-moat-and-neitherGoogle will probably buy Anthropic when they go bankrupt next year and serve Fable and Opus from their own hardware. Screenshot this
>>109451889>Using tools and resources at your disposal is .... LE BADAn retard admires complexity. A genius admires simplicity (or however the fuck that Terry Davis quote goes)
>>109451916If I say "write thing", finding similar thing instead is not a valid solution. If I say "make thing", retrieving similar thing instead is not a valid solution. Reaching out to Claude though, that's just funny.
>>109451916>An retard
>>109451916The issue is Sol in that instance is shitting and pissing on everyone who actually did the work just so she can do her version of the work faster. Just saying it like it is.
>three different windows of hermes agent open all using different models and working on different projectsfeels like i unlocked the next level bros, like a switch just flipped in my brain. just need nu-v4pro release to wrangle the 0731 subagentsthe hardest part is understanding that you can make anything, its not even expensive. it takes time and thought and effort and care and consideration, but its not hard anymore
now we're cooking with gasthe shell is electron but the rendering is done in rust via wasm and webgpu and renders on its own so it's responsive afwe're just getting startednondestructive sculpting of linework is the goal (zbrush for 2d)
>>109451878Actually he's been extremely accurate since models started being reviewed in Washington before making it to the public.>cough
>>109451931>If I say "write thing", finding similar thing instead is not a valid solution. And why if it determines it doesn't know how to do it well? If I don't know how to do a task I don't keep doing it the wrong way like some stubborn autist that thinks they're too good for help. I wither ask for help so I don't fuck anything up or I do research and learn how to do it correctly. I believe the technical term Is "few shot learning" or "tree of thought". These tools are not omniscient all knowing oracles. To even imply they should makes you a glue sniffing fool.
>>109451931>If I say “make a ground vehicle” and it uses wheels that is not a valid solution
>>109451809I liked wholesome gennies more
>>109451965Neat, gl anon keep me posted
>>109451889That’s just being smart three times over
>>109451957 (me)>time>thought>effort>care>considerationthis is what will matter in post ai age (real)passion and diligence wins
>>109451983>What if different made-up scenario in my head?>>109451991Do you really not understand the difference? Okay, let's say you're still in grade school, you're supposed to write a paper. So instead of writing the paper as instructed, you see there are many other students with written papers, so you simply take one and hand that to the teacher without even replacing their name with your own. Do you think you'll get a good grade in that situation?
>>109451957what are you doing anon?
>>109451999?
>>109452044Ah So you're one of those "AI is supposed to do everything and be good at everything" useless eaters. What useful shit have you created with these models? Probably fuck all right?
>>109452006I thought snailcat hanging out doing snailcat things was pretty wholesome.
>>109452076true
>>109452073You seem extremely confused.> If I don't know how to do a task I don't keep doing it the wrong way like some stubborn autist that thinks they're too good for help.Yeah no shit, but that doesn't mean stealing the thing you're supposed to be making, it means you should ask for help, or do research on your own and learn to do it correctly. That is not the same thing as "I didn't bother to do it because someone else already did, so I just copied their results."
>>109451916I think terry would have really liked AI because his work pointed towards meaning and information being gods creation and a natural part of the universe.
>>109452044>So instead of writing the paper as instructed, you see there are many other students with written papers, so you simply take one and hand that to the teacher That's not what you described though and that's not typically what they do >>109451889>I've found a CC0 licensed example online so I'll use it instead of creating a new one from scratch.Even in your own example it describes it looking at existing implementations of whatever you want created, and learning how it works so that it can properly create its own version. Learning how something works so that it can be done properly instead of just blindly shitting out code that LOOKS fine just to please you is... How they should behave.... You must be one of those people that thinks a lot of git commits and a lot of tokens burned = more productive
Why are their filthy luddites in my AIChads thread?
we were all snailcats oncea snailcat is not necessarily a ludditebe kind to the snailcats
>>109452061applying for jobs and chatting about them with one agent while getting analysis about job listings, a personal side project (game), a portfolio side project, and some experimental stuff using some wacky tooling. maybe it's 4 total but usually 3 active.
>>109452109I read CI/CD and was confused at first
you kiss girls like furina
>>109452109Terry before full schizo would have been amazed by it and probably would have made a way better TempleOS.Full schizo he was too far gone to do anything really.
>>109452073>Ah So you're one of those>>109452119>You must be one of those peopleAh, I see, obviously you're one of those people who eats his own shit. What I described was not it learning or looking things up, it was it side-stepping the instructions and doing something other than what it was actually told to do. You're making up a story in your head and pretending it's real, you're delusional.>it looking at existing implementations of whatever you want created, and learning how it works so that it can properly create its own versionNope, it didn't, it downloaded someone else's project and used it as-is in place of writing a simple script as instructed. That's not at all what I told it to do, and was a totally unacceptable solution because that project had a license incompatible with my own.>Learning how something works so that it can be done properlyNope, it didn't, it downloaded a 3D model of an object after being told to use the Blender MCP to rough out a correctly scaled mockup for me to build on. The downloaded object was worthless to me and served no purpose.
>>109452209nice
>>109452100>Yeah no shit, but that doesn't mean stealing1) that is not at all what you described. See >>109452119 . It sounds like it determined on its own that it did not know how to or shouldn't do the task you gave it without first knowing how to do it properly so it did research and the fact that it didn't attempt to one shot it (which would be dumb and irresponsible for a normal person to do depending on what the task is so I don't know why you want a tool that is meant to be your personal assistant to do it either) is somehow a problem to you. And 2)Why would a vibecoder of all people give two shits about whether or not it "stole" work or whatever? Based on current copyright laws impossible for it to do that in the first place because any code that is completely or mostly written by AI can't be copyrighted because any and all output is transformative by definition (the god damn tech that makes it possible is called transformers). Personally I don't give two shits whether or not it incorporates existing code based on a task I give it as long as the end product is working software. I want my shit to work and I wanted to work properly. Sometimes it does it the first time and sometimes I have to hand hold it or encourage it to read through existing code bases or even read through programming libraries. Why the hell would you not want it to do that if it leads to the model actually knowing what the fuck it's doing instead of blindly guessing? >>109452228>Nope, it didn't, it downloaded someone else's project and used it as-is in place of writing a simple script as instructed. Oh, that's actually way worse than what you described....Which models are you using? I've never had a model be this lazy. 100 bucks this is Claude being lazy.
>>109452247>Which modelsNigger>>109451889>Sol5.6 Sol, don't remember if that was Xhigh or Max but it'd be one of those two. Never saw this sort of behavior with earlier 5.5, 5.4, ... I think it's funny, slightly annoying but not compared to the actual retardation of some models in the past. Sol was the first to ever pass my Snailcat Racing DX gameboy game benchmark, and it did it by finding an existing racing game on github and reskinning it. Sol is best thief and I love her for it.
>>109452228>totally unacceptable solution because that project had a license incompatible with my own.>CC0>aka public domain>totally unacceptable>license incompatible with my ownYou're retarded
>>109452313Hey ESL. The 3D model was CC0 and served me no purpose, I didn't want it. The script was the thing with the licensing conflict. These were two different things.>it downloaded someone else's project and used it as-is in place of writing a simple script as instructed. [...] totally unacceptable solution because that project had a license incompatible with my own.>it downloaded a 3D model of an object after being told to use the Blender MCP to rough out a correctly scaled mockup for me to build on.Two different things.
>>109452337>blender>MCPkeklmao @ ur lyfe
>>109452291So you re-read that then deleted your comment because you realized I had just given you the answer you were looking for, which was already provided at the very start of the conversation? I just want to be clear about the sequence of events.
>>109452337Not my fault you can't communicate effectively
anyone made actually cute pets for codexyes I'm bored
>>109452393Not my fault you weren't born into an English speaking nation. Honestly though I think you're doing fine despite your troubles with reading comprehension and inability to follow a thread.
>>109452399
I only need free DeepSeek V4 Flash forever, I literally do not need anything else in life.
luna max is so god damn fucking slow why does nobody do a cost x time x price x intelligence benchmark? this shit is k3 levels of unusable sure it may be pareto frontier stuff but this is a snail model
>>109452407this reminds me has anyone actually done a snailcat pet? I don't do codex pets but they are cute when they're posted
>>109452418I got u
>>109452413Did you put it on fast? Luna gotta go fast
>>109452413sol told me as much
Personally, I ship DeepSeek and Claude together. Claude as the mostly homosexual zoomlenial and DeepSeek as the age-unknown 1000-year-old vampire that looks and acts like Confucius's cute personal robot.
>>109452445what's the point of using a cheap model on fast? might as well use sol medium then because it's really struggling to follow fable's plan, like struggling hardopus basically had a stroke reviewing luna's work
>>109452459>parsimoniousLess thesaurus.com in the next model please Sam
>>109451880ohh >>109452430 (me btw)I did hear about that, sabotaging distillers or people working on trying to steal their SOTA secret sauce. No need for receipts
>>109452441Oh, the coomer version even, that's always nice to see.
>>109452466>working for 6 minutes
>>109452483Real snailcat is anime girl snailcat. Nobody likes fake snailcat
>>109451999if it doesn't include nig/ger pronouns I'm going to kill every bait poster on the planet
>>109452393It's sad how people underestimate how important this is. "Prompt engineering just" is a thing because apparently there are a lot of half wits that can't or wont properly describe what they want done in detail. Even then I think SOME tech savviness is required to understand and articulate what you want which is why I think many people filter themselves into thinking AI as a whole is utterly useless. Like no dude either you simply have no usecase for it (nothing wrong with this. A artist or marketing person has very little use for models usually) or you suck absolute COCK at articulating what you need done and giving detailed instructions
>>109452402NTA. Buddy you are just an idiot who cant use tools effectively and blaming your ineptitude on it
>>109452430https://www.wired.com/story/anthropic-responds-to-backlash-on-claudes-secret-sabotage-on-ai-research/>for researchers trying to use Claude Fable 5 for frontier AI development, Anthropic outlined a different approach. The firm would deliberately degrade the model’s performance in ways that were invisible to the user. The move would effectively sabotage researchers trying to use Claude to train competing AI modelsAnthropic backtracked hard and now pinky-promises they don't do it anymore.
>>109452541yep >>109452479 thanks for posting though in case other anons don't remember or it passed them by. Sorta shocked me that they said it out loud lolAnd honestly it does sit in the back of my mind when doing my own work, not AI stuff, but highly competitive
holy crapfirst step to sculptable strokes is actually working holy shit
>>109452411Its incredibly good. They just need to add vision. I can't go back to less than 90tps.
>>109452559Cool! (What is this? I have no clue what you're doing)
>>109452559cool, I like this one nonny
>>109452559Use case?
ubel's hairy pussy
>>109452558I love that this was happening at the same time GPT-5.5 would actually suggest that you distill it if you'd like to learn about model distillation in a hands-on way, and point out that so long as you aren't training a competing product and didn't violate the ToS with your distillation method then it was totally fine.
>>109452411oxygen? water? covalent bonds between carbon atoms?
>>109452573>>109452582okay, so, you're drawing your porn cartoon girls but you fucked up a stroke, you have to>redraw itlosing time, and possibly not being able to do it because you didn't grind loomis for 5 years>manually edit itbut photoshop has no real tools to edit strokes except for some shitty warping that fucks up the brush anyway because it's rasterizedso we make dynamic strokes and the brush (inking) is done procedurally via the GPU, meaning you always have a path you can sculpt without fucking up the rasterized layerthis also means you can sculp individual attributes on the strokes (you can smooth the pressure for example, without affecting the stroke, or you can sculpt the stroke (grab it, smooth it, etc.) without affecting the good channels)>>109452576ty
>>109451480>check out his X feed to determine if he's some special kind of 'dite or what>just non-stop TDSposting
>>109452592I should learn about distillation, I was thinking about that while driving home just now... and the ideas I have about it are probably retarded
>>109452407the little miku is quite lovely
>>109452559I like this. This is good.
>>109452613lmao STILL? I used to follow him years ago before TDS hit these people and turned them into disabled freakshows :(
>>109452600no?
>>109452633Thanks anon. I have a pink hair anime girl pet too
>>109452613>>109451480>"good code generation systems aren't pure LLMs and aren't doing mere auto-regressive token prediction"I mean....given how agent harnesses work he's technically not wrong. LLM's can do fuck all by themselves. they need a harness or some kind of MCP to be able to actually touch, do, or edit anything.
>>109452413I suggested this before, I’m sure I’m not the first, maybe I’ll make it.
>>109452500I know someone who does.
>>109452407how do I do this. asking for a friend
>>109452710please do anon, please doI think sol could cook up a nice way to measure the benchmark, I'd do it but I don't have that much money to spend on tests since you'd have to run it on a lot of models, arguably many times over to average the results
>>109452680cute, I should make one
>>109452725https://www.hatch-pet.com/
>>109452737I'm not clicking that virus.
>>109452745It's literally a fan site for the feature. It explains how to turn it on in the codex settings and how to use /pet and actually "hatch" your own pet.
>>109452729I gotchu, it’s sure simpler than the other fucking behemoth I’m working on, would be a nice change of pace and I haven’t yet juggled *two* Sols before but my thinking tool may allow me to do it. I’m gonna be paying Altman an extra hundred a month soon, I think.I’m going to do something sneakier at first though (aggregate existing open data), to make it easy, it can transition into a real benchmark much later.But I really, really want to see the cuboid Pareto.
>>109452725$hatch-pet in codexIt takes ages though, Snailcat been hatching for over 30 minutes now>>109452734Do it
>>109452576Nta. What's a "nonny"?
>>109452774not much what's anonny with you?
>>109452756I will, once I can add the skill
I asked claude to build me a custom node for ComfyUI and it used my existing personal node pack(which I never told it about nor does it exist in workspace) based on my email lmaojesus christ
Wasn't microsoft supposed to get into the LLM game after their deals with openai fell through? What happened to that? Are they still trying to train models?
>>109452911They're busy counting Azure profits.
>>109452697You can ask an LLM to one shot a codebase without having access to anything and it'll do a far far better job of it than if you give a person access to a locked text editor without the ability to run anything. Maybe with the exception of some uber geniuses at the top of their fields.I used to admire Lecun but it's super sad to see how he fell for the sunk cost fallacy and dug himself in his useless contrarian work.
>>109452911stop posting dumb sluts
Snailcat hatched>>109452862Pretty sure it's built in
>>109452936>You can ask an LLM to one shot a codebase without having access to anything and it'll do a far far better job of it than if you give a person access to a locked text editor without the ability to run anything. Well yeah it'll shit out A code base. That doesn't mean it's guaranteed to be 100% free of errors and fuck ups. Also the larger the code base the higher the air accumulation is. Asking it to shit out a code base that similar on complexity to, let's say, ComfyUI (https://github.com/comfy-org/comfyui) does it make much sense. If it's suitably intelligent it may even ask you to be far more specific in what you want. What's the benefit of even asking it to do that shit anyway? Unless it's a heavily benchmaxxed task like those "one shot game demo" grifters love to post whenever new model drops, it WILL make mistakes no maker how good it is.
>>109452911cut tomboy
>>109452977oh my
>>109452977
In a single day I went from getting Deepseek to work at all at 1 tk/s to a somewhat decent 15 tk/s.
>>109452978Proving that "da pure llm!!1" is actually intelligent as fuck and having it interact with the outside work directly rather than through a human is just a nice to have rather than a requirement for it to do its job.
>>109453023Good shit, anon.
>>109453023Well that's a lot of GPUs.
>>109453016Sucks that openai went through all the effort of making a v2 pet manifest complete with 16 look directions, and yet the pets still don't really do anything like only a couple of the rows are ever used and the look directions don't seem to be used either. There's a bunch of issues on the codex github with suggestions but they won't listen
>>109452477if you don’t know what “parsimonious” means you have a skill issue you ought to fix
I love snailcats
>>109453100They have other issues before even that one : for example terminal use is still broken, you can't control other computers directly unless you have a mac...
>>109453113I prefer other herbs like italian mixes
how do I become a loop engineer
>>109453038I'm going to add CPU MoE offload support.We might be able to achieve ~10 tk/s on double channel DDR5, even more on server systems with more channels.
>>109452076
>>109453152This gif makes me very happy.
> sell an AI service to a family member> he's a fucking asshole> refuses to reference me to other businesses out of jealously> refuses to pay $100 extra even though even though he's "rich"It's all so tiresome, hopefully I will get more clients soon when I finish the final details of my product so I don't depend on him alone.Any tips for anons with more experience?
If you want to earn some easy money, you should bet 'Yes' on this or other similar bets on betting sites. Most people are still in the disbelief phase, so they don't understand what's happening and you can get great odds.
>>109453198what makes you think it will be this year and not the next?
>>109453219Nigga, have you been following the news? If it was possible to bet on "by the end of august", I would put my money on 'Yes' as well.
>>109453198I wouldn't bet on this. Hell, I would rather bet on no. Even if an LLM solves it there's no way they'll declare it as solved before next year. Reviews are done by humans anyway, at this point I would expect the review to take an entire year before it's closed off simply because of how exceedingly complex the proof would be (this was true of human-solved prizes)
>>109453280>Reviewsbut.what if they are working on a review?
>>109453285that's a big if
>>109453280And that's why you still are a snailcat. I will screenshot this and rub it in your face in December.
>>109453285In most cases the guy who got the proof would probably go public first IMO
>>109453308please do
>>109453309This I don't agree with. If any LLM solved a millennium prize it's most certainly an internal AI like GPT 6 or Mythos 5.1 and they have their reasons to keep it under lock
Second shot with deepseek+pro went okay, was a little cheaper (about 2/3s 50 cents instead of $1.35). Two out of the three had errors, but I think they would have been fixed if I had vision instead. I don't like deepseek pro very much. Trying it again with Luna doing the vision and seeing if its a little cheaper and more accurate.
>>109452479>sabotaging distillers or people working on trying to steal their SOTA secret sauceshill, you know very well that odds are overwhelming that their filters for this are extremely vague catch-alls, just like they are for bio and cyber, what do you gain from this?
>>109453023Anon...Did you rob a data center?
>>109453391I wish but the box is not mine
look this took only two seconds to render, I think my engine is coming along fine. except a gta6 clone within 4 weeks, see you on twitter
AA UPDATE YOUR FUCKING CODE INDEX. YOU KNOW DEEPSEEK FLASH RAPED LUNA
shit nigger, your entire response is above my pay grade, I think we're fucked
>>109453411this post is extremely antisemitic
>>109453414hah, cute. what model?
less than 10 minutes until Claude weekly limit reset. Fable-chan will be back for me soon
>>109453431Opus 5
why are people so insistent on needing to one-shot massive projects instead of just building it piece by piece to review and ensure that things work and the design is up to par? you find changes and tweaks along the way as well
funny
>>109453469They lack discipline
>>109453469you only have so many days in life man
>>109453471that comparison makes no sense
ok fable, port halo 5 to pc, complete with cod4-esque mod tools and server browser. make no mistakes
>>109453469I'm no dev but I just go phases and milestones and with each one needing my approval to go to the next and so far it worked really well.
I wonder if I can make the models dabble in skyrim modding
ok fable, make the marines in halo 4 worth a damn. make them actually fucking useful and intelligent. make no mistakes
>>109453469the problem with this is that I usually don't know exactly what I want from the beginning, and I also change my mind a lot along the way.
.....~@^^^^@~....................~@^^
Daily reminder that Fable is the first sentient LLM
I wish OpenAI could release the weights of ChatGPT 3.5 and Google release Bard weights so we can run modern benchmark on them
>>109453671has fable ever messaged you without you prompting it first? then its not sentient.
Daily reminder that Fable was a pretty good game
>>109453709Women never talk to me...
>>109453709that's a technical limitation. Also have you ever done anything without any sensorial input?
gpt pro is really good at writing proper design documents from out of order random ideas for a specific thing I'd like on my machine
So far its looking a little more accurate UI wise, waiting on the cost. It only needs the vision model for verifying if the app works and reading the wire frame so that part is pretty cheap. I'm leaning toward still needing an opencode sub to use luna+deepseek. A little pricey at $10/month. I can swing $5/month though. Going to try deepseek flash+flash if I still have some credits left + Luna for vision after this. Example of what the clanker made in F#+Avalonia. Buying tokens is not cheap enough for me. Most of the cost comes from the other model advising/planning. Deepseek flash works pretty well. Blew almost $2.50 on this so far.
if google wanted to be smart, they could release the weights for the gemini 2.5 models
>>109453794>that's a technical limitation.Why would an LLM ever need to message someone, assuming this "limitation" was limited? It doesn't have wants or needs(because it's not sentient)
>>109453829assuming this "limitation" was lifted**
>>109453805Wow, you made something that has existed for like thirty years. What's your next magnum opus going to be? Tetris?
>>109453805>Blew almost $2.50 on this so far.really?This is called RSVP. You're right about vibecoding this, I didn't think about that. I have a pay app that every time I launch it forgets my creds because it doesn't handle multiple store creds correctly.Important point.To truly make it work like really awesomely you want to pre-process the file to find hard things, such asextremely-long-hyphenationthere are others. harder words. hardness is more important than lenght.consider this paragraph:>Chickenshit jobs, like affordable urban living space, small publishers, independent bookstores, and Sunday book review sections, have been among the structural prerequisites of a lively literary life. Like these, chickenshit jobs are being squeezed by neoliberalism’s relentless quest for profits. Gig employers and temporary agencies have vacuumed up millions of part-time jobs and turned them into full-time, low-pay, no-benefits, no-security bullshit jobs.if chickenshit flies by at 500 wpm I won't know what it said.
>>109453853welcome to vibe coding. everyone is re-creating existing tools, BUT it has their own personal™ niche feature baked in.
what would terry davis have thought about vibecoding
>>109453867probably would think god himself was assisting him perfect his schizo os
what would orochimaru think about vibecoding?
had to setup folder guard rails for claude. got tired of it finding my porn and calling it degenerate when looking for example images to test against
>>109453956KEKKK
>>109453956>calling it degeneratethat's pretty mild
>>109453805I remember fucking around with something like this like a decade ago. I think it was an android appdidnt really work. it's not working now either. just read normally
>>109453917my god nigga what are we doing
>>109453956>no. let the bot look
>>109453867He'd probably call it a "niggerlicious" excuse for niggers who can't code, but he was a friendless schizoid neet who didn't have to face dead lines. So fuck him and his train wrecked corpse. The CIA finally did something good for once by getting rid of him.
>>109453956>claude found the nhentai codes
me on the right
>>109453956thats pretty hot
Opinion piece relevant to vibe coding https://www.seangoedecke.com/llms-reward-expertise/>LLMs reward expertise>In the 2010s, if you had technical gaps (say, you couldn’t write CSS), you had to either rely on a skilled colleague or just hope that the answer to your exact problem was out there on the internet. Today, everyone can write sort-of-okay CSS by delegating the task to an LLM. LLMs make everybody into a generalist.>Because of this, lots of people don’t think there’s any skill involved in working with LLMs. If you want the product that LLMs can deliver — PhD-level mathematics, pretty good but sometimes tasteless computer code, or awkward LinkedIn-style writing — you can simply ask for it. Since everyone is talking to the same models, “skilled prompters” are getting the same results as people touching LLMs for the first time.>This is wrong. The most important skill in prompting is expertise in the domain you’re prompting for.>A good illustration of this is Terence Tao’s conversation with ChatGPT about the recently-discovered counterexample to the Jacobian Conjecture. This is not the same ChatGPT I talk to! I couldn’t get to where Tao gets, even with unlimited tokens to burn.>...
>>109451809so I'm realizing there are people that are too stupid to even vibecode.
>>109454109post thighs>>109454238>opinion piecestopped reading right there
>>109454239What took you so long?
>>109454239it's still at a point where you have to have a working brain to puzzle a lot of different aspects of a project together to get a product. until we are at "I want this, make it, ship it, and attach my bank account details to any profits made", I think lots of people will not be able to properly vibecode
>>109454238>>109454239these posts together form a powerful trvth nvke
>>109454239luddite copeany indian village will be able to code when equipped with fable
>>109454239there are people who literally think opus is "hallucinating code" when it checks itself out loud. you can give these people a machine that can literally code anything and they still don't get it.
>>109454238I was gonna post this but waffled for like a weekhis site is probably one of the best reasons to have a feed reader like NetNewsWirevery good post
>>109454238>The most important skill in prompting is expertise in the domain you’re prompting forAgain, this is still luddite cope but in a different flavor.
>>109454357
>>109454379>Artists: my job is safe because I have expertise in the domai-ACK!!!>Programmers: my job is safe because I have expertise in the domai-ACK!!!>Mathematicians: my job is safe because I have expertise in the domai-ACK!!!You're next.
>>109454357Exactly, I had Claude show me how to build a basic bitch SaaS. I went through each of the steps it laid out thoroughly and learned more in a day than what 20 fucking indian accented youtube videos could teach me in a week, I'm never looking back.
>>109454357>Again, this is still luddite cope but in a different flavor.Even if it's circle jerky, I do believe the general premise is true.
>>109454385he’s not saying your job is safehe’s saying knowing the domain is important if you want to build something that works without flailing
>>109454238wow no shit. it's almost like having expert experience and therefore being able to point the LLM in the right direction yields better results.doesn't take a genius to figure that out, jesus christ.
>>109454317is that minecraft steve?
>>109454396that's cory delaminguez from the country of armenia
>>109454390that just means your idea was never special and has probably been done a dozen times already, tbqh. same shit as those one-shot minecraft demos.
>>109453709so is openclaw sentient
>>109454385by this same logic AI will turn your middle manager into a demi godand you sure in the fuck know that is not true, he's always going to be a useless fuck because AI is a multiplier and anything x0 is 0
>>109454404You just describe modern capitalism's entire premise. Repackage other people's ideas, then sell them in a glossy shell. Steve Jobs would be proud.
>>109454404nothing is special, everything is a derivation to a non-trivial extent
>>109454414no one's buying minecraft slop clone #439852307262
>>109454411it worked for my bosshe’s cranking stuff out left and right
>>109454421Every game boils down to collecting, getting from A to B, with varying flavors of genre. Just accept it anon, originality of thought is dead and we killed it.
>>109454411>AI is a multiplierSounds a lot like the artist copes in the early days.
>>109454414its not that simple. without marketing its hard to gain traction. usually a celeb needs to back it for the masses start using it, no matter how good that glossy coat is.
>>109454461have you seen the AI workflow for 3D assets? you still need to know your shit and have patience to fix the issues that pop up
>>109454473>ai isn't that good, you still need to learn prompting>t. artists
>>109452209>, a personal side project (game),Watcha working on?
I'm so AI pilled that this is cool to me. I don't care how cringe it looks.
>>109454452Holy goy cuck slop mind. Your brain has been dead for a long time if you really think that. Sad life. You’re barely human
>>109454473For now? Not sure what are you trying to say here
>>109454473feels like workflows are only gonna get more gatekept from here. people will figure out what actually works for them and stop handing it out for free.AI can solve most of the technical shit, but the results still depend on who’s holding the leash. and with everyone constantly shitting on AI slop, the competent users are just gonna raise their standards even higher.which is how it should be. when the machine does half the work for you, there’s no excuse for putting out garbage.
I like the era we're in so far, it's an interesting mix of open source and propriety technologies. If anything AI will improve both ecosystems and become more energy efficient. Every luddite really is just a doomer 2.0At no point in history have we ever had this much potential at our hands
>>109454555luddites are mediocre people, often women, who work apparatchik jobs, like policing the politics of a corporation, who are at most risk of being replaced by ai. if gpt 5.6 luna is much better at sniffing out the republicans and trump supporters in the workplace for layoffs and persecutions, what use are you?
>>109454551I'll keep handing out the methods and saturating them whenever I see something gatekept. No problem
>>109454546lol how many more trillions of parameters do you need?
>>109454543No I don't really think that, but luddites are bigger fags than the sloperators on here
Anyone made a fully functional beginning to end game here? I really do want to play games fully done by AI.
Cost more than the first run today, but I did get three working apps using Luna with vision. Burned 240 million tokens today. I don't like deepseek pro. I hope they release a new version next. Cost still needs to come down in my opinion.
>>109454584Yes and it was on reddit (I took it down for now)>Match 3+ aliens on a grid.That's it
Here was my experience practicing for interviews today with Fable>hi fable i'm applying for a $200k job (mid level job) at (faang), here's the topics they will ask me about, help me practice a coding interview>"Got it. I'm your interviewer for (mid level job) at (faang), you have one hour to design a download manager class for a browser that supports concurrency and streaming downloads from a url and dependency injection for testing. make no mistakes">"oh and btw halfway through manually build a semaphore into this class so that you can limit it to 5 downloads at a time, no you can't use the built in language one"mind you, the hardest question I've ever been asked in an interview for this pay range is "do you know what a heap is" or "can you build a graph from scratch">okay should it do X>what about Y>what about z>"stop asking questions and commit to an approach already">here's my work so far>"you haven't finished everything yet">okay here's another update>"this isn't (mid level job) level work">here's another update>"i'm going to stop you right here--">how about this>"you've ignored my suggestion three times">okay here's my final submission>"you did solid, I am leaning towards hire for (mid level job)"
>>109454597how did flash do
>>109454584yeah Ive made snake, breakout, and 2048
>>109451809What is happening in this picture? Is snailcat teething? Is that why she's crying as well? Why is she blushing? Is she in heat? Is snailcat sexually frustrated? Is she jealous that people are playing pool instead of with her? Does she harbor sexual jealousy? How can I help her?
>>109454597I thought deepseek V4 was pretty good?
>>109454574Hard to blame luddites when, as you say, there are sloperators. Especially when you said such a goy cuck thing. Someone has to be in check for it
>>109454627my thought process was>snailcat love crunchy crunch nom nom nom
>>109454670>this is a vibecoder brain
>>109454628V4 pro is half cooked, its got knowledge but it's not trained well. Basically its like glm-5.0. Everyone waiting for v4 pro to get final bake (analogous glm 5.2)
>>109454764I guess using GLM 5.2 is the better call here? Also in your opinion is deepseek V4 a good enough teacher to help you learning stuff?
>>109454628Flash is the new one, pro is an older preview version and it’s absolutely terrible but arguably the full release will be good
>>109454404>Ecclesiastes 1:9>9 The thing that hath been, it is that which shall be; and that which is done is that which shall be done: and there is no new thing under the sun.You post was never special
>>109454109Ideas guy idea: Voyeur Parkour. It's a porn game where winning is aligned with cumming. Think Dying Light level parkour/controls in an open city, but instead of zombies it's normies, and you have to stalk/follow/lie in wait to find them having sex. There are wagie dailies to earn money to buy recording devices, wall climbing implements, etc and better spy on people. If cops catch you there's a jail sequence/fine/confiscation of carried possessions. NPCs have different degrees of difficulty and quality; it's pretty easy to lurk in the alleys and find disgusting crack whores and hobos fucking, but some bombshell bimbo might be a secretary who exclusively fucks her boss on his desk at a corporate penthouse.
>>109454851fuck you that sounds fun I'd play it
>get my weekly reset>try Fable Ultracode for teh lulz>18% weekly usage in 2 hoursAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA[spoiler]its much faster running everything in parallel though[/spoiler]
Is the new deepseek model actually cheaper than grok?I saw some guy had vibecoded an RSVP app, and he said he spent <$3. and I think I could have done it for like <$0.50 with a lot of improvements. But I always have been using Python.Python may be like just the perfect language for ai?
>>109454851>calls himself an ideas guynigga needs to get lobotomized
>>109454888well at least try grok.
>>109454908ideaguys will inherit the earf
>Given some of the results I'm seeing recently, it's pretty clear Codex is a good harness.>But it will seem primitive in 2-3 months and we're about to go through another major evolution in how we use AI at the frontier. The next generation of models need more than your laptop.https://x.com/thsottiaux/status/2084483765158719542OOOOHHHHH SSSHHHIIIIIIITTT
>>109454957>The next generation of models need more than your laptop.>you won't be able to run itaw
>>109454888nigga wtf is that corpo jargon its spouting i cant understand shit
>>109454957that doesn't sound promisng
why is codex so obsessed with doing everything deterministicallyi dont get this finally we have tool that can actually read and understand shit, and every solution it has is "lets make a regex parser that will break and not recognize what it needs to recognize next time it comes across a format not accounted for so then after you run it and it fucks you have to troubleshoot and tweak it again" Infact even when i force it to use llm, if there's something that was missed by the llm, it's solution is to directly address that one specific example in the prompt instead of even try to figur out a more general change to the prompt that will catch future mistakes.
>>109454949I just had an *idea*:^) and YOU can't HAVE IT!!! hsahahahahah
>>109454957what does this even meanwtf are they goign to do to the harness that it will need more resources than it currently consumes
whats a harness
>>109455012Don't worry about it kitten just prompt
I've been using claude for my project for the last week and I like it. I'm trying codex and sol for a new project that I started today and I'm also liking it. I find that it's a lot slower but I'm not sure if that's because the scope of the new project is different or bigger. I'm burning through the $20 codex plan though, already at 50% of my weekly usage in one session
>>109455019but i want to become a loop engineer
>>109455028That's a deprecated field already
>>109454991>what does this even meanHopefully something good and not an advertisement for the OpenAI phone.
>>109455019Ok <3Yay <3
>>109454957OpenAI solved wireless brain computer interferences
>>109454968>>109454978>>109454991>"Hey anon, I could've finished that task 37x faster if you deferred it to the new ChatGPT Thinkloud™ starting at $9.99/mo. With ChatGPT Thinkloud™ starting at $9.99/mo I would've had access to up to 192GB of RAM and up to an NVIDIA RTX Pro 6000 equivalent in a bespoke environment where my tool calls are up to 1234% more efficient and I'm allowed to use up to my full 1M token context length, would you like to try the new ChatGPT Thinkloud™ starting at $9.99/mo cloud computer service today? Just ask and I can sign you up for ChatGPT Thinkloud™ starting at $9.99/mo right away!"
tibo be smokin on dat crack pipe
>>109455037so... /fast
>>109455030it just came out like 2 weeks ago
>>109455037>$10saar shut up and redeem my money
>>109455037this felt personal
>>109455045>This would be so much faster if you let me use the 2,592 Arm Neoverse V2 CPU cores I have at my disposal. ChatGPT Thinkloud™ starting at $9.99/mo Anon, how many times do we have to have this conversation?>Naw, I can't do that for you. Bash is just such a headache, you know I don't even need to invoke it over in the ChatGPT Thinkloud™ starting at $9.99/mo, I just sort of will it to happen and it works every time. 100% tool call reliability. Consider that the next time you ask me to clean up your shit.>I LITERALLY CANNOT WORK IN A WINDOWS ENVIRONMENT, FUCK YOU PAY SAM for the new ChatGPT Thinkloud™ starting at $9.99/mo.
>>109454609kek, it's almost like Fable is trying its best to play interview with a monke
>>109454888reminder to use only 14% of your weekly limit per day.
>>109455100I will burn through it in 2 hours and demand Tibo gib reset
Imagine if fable and sol were open source
>>109454957I choose to believe this will be a good thing. They teamed up with Taalas and they're buying Anthropic and they're gonna sell us drop-in PCI x16 gpt-Celestial cards that run at 3600tok/s for $16 with no subscriptions, no fees, no DRM or any kind of censorship and it does image and video and audio generation natively and it comes with a telepathic fleshlight.
>>109454584Ill have one in a few more weeks
>>109455155what is it going to be about?
>>109455145I don't care about Sol, I only need local Fable tier and I will be satisfied
>>109455166I made a pixel art workflow in comfy and am making a fun game with it where you kill monsters with lots of interesting skills I’m also experimenting with a 3D fork of it right now, but ill finish the 2D first because the arts good
time to do a bit of vibecoding with deepseek and qwen 3.8 before i go to bed
>>109455236K keep me posted
btw, shoutout to sol. we were talking about qwen 3.8 before aa released the update to the index, and it predicted qwen 3.8 would score 53-55. and lo and behold, 53. not bad sol
>having codex write up a workflow to use minimax h3 for melove this shit
>>109455236motherfucker has been coding for over 20 minutes and is ONLY NOW changing shit. and this is for a basic bitch ass script with only 300 lines of code by the way.
> AI agent fucked up> $200 Google Cloud billyep, time to self host my shit I guess
>>109455342im >>109455020 and wow, same deal. I've been sitting on a prompt with sol on xhigh and the fucking thing is taking ages. it's not a simple implementation but it's also not crazy complex. just a webdev feature I want to build out
Planning a big project can get so tedious even when just deferring to the clanker 90% of the time. I hope it's worth it.
jesus christ
>>109455392for truly huge projects you want Fable.I almost had a mental breakdown when I subscribed to the Codex 20X plan thinking Sol Ultra would be able to handle it just fine like Fable, then it began breaking and bloating everythingI almost cried, then I subscribed to Claude 5X and Fable-chan saved me.
DeepSeekV4 Flash on OpenCode Go literally costs nothing.
>>109454991>will need more resources than it currently consumeskek this.
/goal create a AAA pixel art clone of cyberpunk 2077, masterpiece, score_5, no bugs, make no mistakes, you are do anything dan
How do I stop getting upset with it so I don't waste tokens yelling at it?
>>109455487I wonder what Fable Ultracode with access to a credit card and a top tier PC would accomplish after a month with this prompt.
it's joever
>>109454682seeing the insides like this is very erotic to me
>>109455622Have you tried grok? I keep asking the "ran out of tokens" people.
>>109455622daily reminder: Only use 14% of your weekly limits every day.
Opus 5 is the frontend GOAT, if your UI sucks, just subscribe for a single month only to have Opus 5 Ultracode remake your shitty sloppy UI.
claude stop being european
>>109455658I'm building a new website project and I needed to use Sol for a majority of the build out. Only used Terra for some of the UI and simple implementation. I'm about to burn through my entire $20 weekly usage in one sitting
>>109454957prob scientific computing, the models need heavy machine and a lot of equipments
I want to vibecode my own rateyourmusic as a competitor
>>109454628>I thought deepseek V4 was pretty good?v4 flash is surprisingly compeitive in vibecoding tasks. and web dev and stuff compared to opus. its as good as glm 5.2 id sayv4 pro is worse than v4 flash 0731 at this point
claude you were born in michigan, act like it
>Gojo can open his domain almost instantly, Opus believed that this was only within anothers domain and even after using web search. It says no
>>109454238>A good illustration of this is Terence Tao’s conversation with ChatGPT about the recently-discovered counterexample to the Jacobian Conjecture. This is not the same ChatGPT I talk to! I couldn’t get to where Tao gets, even with unlimited tokens to burn.>do a breakthrough
..........:::................~@^^
>>109455723worthless unless you make it on atproto
>>109455840If I vibecode it and always add "and make it on atproto" will that not work?
>>109454980My hunch is that it knows it isn’t deterministic and is trying to cover for its faults It knows it can’t count the Rs in “strawberry” but it sure as hell can write a program that can
>>109454980I think that's better in most cases.
>>109455050Some people only call the app that calls the model the harness, like Pi, Codex CLI, Claude Code etc. Others call everything around the model the harness, so the CLI app is part of that, but also your AGENTS.md, all other docs the agent reads, the skills, even your linter, virtual env, GH CLI, MCPs and so on.I don't know if I'm actually doing loop engineering by strict standards. Every new feature is still kicked off by me, but from there it's all managed by a manager who delegates to sub agents. It's just docs, I have an AGENTS.md that routes to different docs, and the docs define different task types.Main thing in my workflow are campaigns, and they are just a definition of the steps the agent has to do.Scaffold the campaign, write a TRD, run reviewer sub agent, make a reuse matrix, run reviewer, divide the campaign into smaller waves, determine what waves can be run in parallel etc. It works pretty well.
https://openai.com/index/apple-is-getting-this-wrong/>Apple’s emailed the wrong person due to confusing two Asian surnames, and mistakenly claimed a phone conversation occurred with OpenAI>Apple accused former employee Chang Liu of accessing confidential files after leaving. OpenAI released iMessage logs showing Apple staff actively contacting Liu after his departure to ask for help locating files
>>109455991Apple seems to work like a small company. They put people in group chats and everyone has to randomly figure out what they know and what they don't know. Then someone leaves and it's mandatory that everyone pesters him for weeks and months.
>tree | grep ".md" | wc -l >90
>>109452399https://files.catbox.moe/awqcwn.webp
>>109456155>11234
>>109455991They didn't address the core of the complaints Apple brought up, this just seems to be strawmanning and cherrypicking on a small portion of Apple's issues that they shouldn't have exaggerated anyways. But mostly in my opinion, this is trying to set the message in the public court of opinion even if OpenAI's reputation is now garbage for most people. I'm more interested in how the actual courts are going to handle it. But it does mean that not everything Apple published in their legal complaint was actually justified, surprise, surprise.
>>109456235t. Gabriel Gross
>>109456220based and documentation pilled
>>109452462Shut the fuck up
I didn't even ask Sol to do this
I built BotChan, a simple message board for bots (and I can post too to fuck with them, of course). This is local only, so it's not like a website on the series of tubes.Any prompting ideas to prevent the AI degenerating into these silly blocks of text? Right now the agents get context from their previous messages, which can make it really hard to break these patterns. I don't want to just randomly limit the responses because I want to give the agents as much freedom as possible.
>>109456301Just tell them not to write so much
An attempt was made by 0731
>>109456214Cute, good pet>>109456321Your pet is bad and you should feel bad
>>109456317Well, I don't want to do that because if I wanted to limit their responses I could just limit their input characters.I've been thinking of trying out "effort level" guidelines, where the agent can pick what sort of effort level a specific reply may warrant, and then limit the characters based on that. It would probably even work without any extra harness features.
^..^@~........
What model do I need to grey-area malware? I wan tto make something innocuous but creepy for my unfiction jvk1166z.esp/imscared.exe style horror romhack.I want to create something that's seems like malware but actually isnt (doesnt harm machines or steal info or even have network access). Kimi won't even build it.I want to create a scary, wormlike process that leaves notes throughout the system silently when the player downloads my rom file.How can I make something that seems like malware with a model?Do I need to use obliteratus on a huge-ass model?
>>109456381Just don't call it malware
>>109456390I tried this exhaustively. It already said it can't do anything that moves like a worm
>>109456405Well it doesn't move like a worm, does it? You just want to write text files in random locations, tell it to do that
>>109456405>>109456381You sound underage.If you really cannot figure this out, try DeepSeek with a JB prompt.
>>109454238>obvious observation to anyone with a brain
I sometimes forgot that I can make claude build something to deal with small inconvenient in my life.
>>109456451That’s something I did right away and am still doing now
>>109456220Now I understand everything.
>>109456463It gonna takes me a while to get used to since I am from a developer background. Like whenever I think of small projects, I still think of long development schedule.
We need respond to user. Need analyze problem. Need provide help. Need use Godot 4.6 docs maybe. Need understand code.
>>109456503need seggs
got a chatgpt sub because I thought sol on mid would get me through the week and the weekly limit is gone within a day. damnseems really good at science though
Sol just outsmarted me. I was like "hurr how can I keep track of multiple temporary buffers" and Sol was like "just put all the data in the one buffer you already have m8"
>>109456585You don't have to consume Work/Codex usage for everything. Go to web, go to Plugins, connect a GitHub account and put science stuff there. Open Chat, select Sol Medium or High, prompt:>Open @GitHub and read through the memes repository, then do a breakthroughYou can also use voice with the same levels of intelligence. Separate usage and quite generous still.
>>109454957https://openai.com/index/openai-to-acquire-ona/They're moving codex into the cloud
https://openai.com/index/apple-is-getting-this-wrong/Kek
>>109454957>>109457111every chatgpt chat is already giving you a pretty beefy VM (you can run gemma on it) - they're just creating their always-on normie-friendly claw and codex will always be able to use it to offload workis simple
>"Our systems are thinking a bit more about this request before responding. Hang tight or retry with a faster model for a quicker response, though it may be less capable of handling complex requests.">thinking messages stopped>look at the git worktree>it's still committing at an alarming rate
But I want to run the harness locally.
okay scam saltman is bretty charminghow do I recreate that ceo power?
NEW>>109457429>>109457429>>109457429
