A general for vibe coding, agentic engineering, coding agents, and shipping code with LLMs.## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/## Related generals>>>/g/lmg/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://claude.com/product/claude-codehttps://developers.openai.com/codex/cli## Best free tierhttps://antigravity.google/pricinghttps://opencode.ai/ (free models)## Will there be a codex reset?https://codex-resets.com/## What we’ve donehttps://vcg.gitgud.site (never gets updated)## Previous thread>>110042701
what a dumbass
>>110045789sol is better at turning a planning session into a detailed implementation planreview as you saidit's surprisingly good at digging really deep when searching for bugs, it found more than even fable for me and it's way cheaper. it being kinda slow I just tell it to comb through the code and mark any bugs and issues and go to sleep
Trump requested this. It's mine now.
Guys… am I in danger. We have this pain in the ass report that has like 30 pieces of input. It used to be a pain in the ass project. Took 2/3 daysNow with a python script I got it done in two hours. Which included my 1 hr of review. Im scared to send this over to my boss, but I looked through it and it correct.
>>110045841>in order to smoothen the demand curvesooo openai would like to have the demand be more predictable?how about MAKING IT SO THAT RESETS ARE PREDICTABLE OR DON'T CHANGE YOUR WEEKLY DATE INSTEAD OF PUSHING PEOPLE TO BURN ASAP BECAUSE WHO THE FUCK KNOWS WHAT WILL HAPPEN???What use is an increased limit when I can't plan worth a damn? Sure seems like the users have the exact same problem.Of course double usage is better than no weekly allowance - in a world where all resets are banked resets.
can you faggots make your own general to talk about the twitter account of an OpenAI employee?
>>110045859Chill for two days. Overstate your involvement. If you want to impress him/he's expecting results from the AI, chill for one day, say it still needed a decent amount of human involvement. DO NOT tell him it was entirely automated.
>>110045871I think OpenAI is running a program on unempathetic people where they turn them into what basically amounts to simps for tibo, like imagine fangirls at a concert, that's what those people are for tibo. they want to suck tibo's dick that bad
how many tokens do you get for a sloppy toppy? I need to run the numbers before deciding how to proceed
did they actually implement the 500k context? I updooted to current macos chatgpt version but it still seems to compact around 300k
You can definitely say it's one hell of a humiliation ritual
>>110045870Hot take. I dont think openAI knows what its doing and screwed itself by prioritising high normie adoption
can you guys vibecode me a up to date version of webm for retards? please and thank you.writing parameters is too hard for me, so I need a GUI for making video clips
>>110045914>did they actually implement the 500k context?lol
>>110045870truth nuke
>>110045925I'm actually working on one for myself because I'm sick of using the CLII may release it to github
>>110045914>>110045928tibo trannies blown the fuck out
>>110045925I vibecoded a video recorder that after recording lets you cut out bits of the recording and you can tell it how big a file you want and it automatically messes with parameters to make a soundless webm of said file size. Never having to deal with the (admittedly minor) hassle is nice
If you have ever written "dropping gems, king" under a twitter post in your entire fucking life please come forward at once
>BroetryLMAO
>>110045925you can upload mp4. now more libxvc19492 or whatever just leave the audio out
>>110045943>>110045948really? I like how to the point and simple the interface is on the software. the additional tools of cropping, resizing, rate etc. are really good when you quickly need to share something of a clip you have recorded
>>110045969NTA but it's still a major pain in the ass>can't have audio tracks>4 MB size limitso you end up fiddling with the parameters in ffmpeg anyway
>>110045870>openai can't predict the demand>users can't predict the supply>the cause for both issues is the sameIt amuses me that their attempts at optimizing usage is what caused this ballache in the first place
biweekly check in whos winning and why
>>110045841Hang on. Why are they doing twitter poles for stuff like this anyways? Not all their users can vote and on the flip side any tard with a xitter account can vote even if they never used chatGPT. Dario himself could literally have gone and voted for the one he thinks will hurt you more
>>110045974PS. I'd use the software mostly for sharing elsewhere than just here, but of course you could just include/exclude audio with the settings etc
>>110046002>winningMy grandpa fucked off for 3 weeks long vacations in portugal. He barely uses any tech.>losingPeople who waste their time on tech
>>110045925give me the exact gui and my clanker will clank it
>>110046018I've been to a royal caribbean full of boomersI'm not sure I'd call that winning
>tell opus to spawn haiku agents to do the implementation>come back to find it fixing a hundred test casesTest slop is the new slop innit?
>>110046022like so?
>>110045988the solution causes the problem is like 98% of all problems
>>110046061if you want the same functionality why not just use webm for retards as is?
it doesn't matter how many times I try it, sol's absolute autism when building UIs never fails to crack me up
>>110046069It's not that bad honestly, bit of refinement here and there and it would look quite usable. I think it's better than what Claude output at default
Anyway, here's a new video. Populations with settlements + road networks overlaid on different maps>>110046069It does work though, can't complain about it. That's the usual pattern with Sol, absolutely abysmal UI, really good backend.>>110046078I don't know, I absolute adore Claude's taste
>>110046061I'll give it to the clanker but remember, any elements that aren't shown are going to be decided by the clanker.
>>110046040there was a paper that showed that pretty much all frontier models would score higher in programming benchmarks if you included into the prompt "no tests"maybe there's something about it
>>110046069I don't understand people bitching about design for simple tools like it. The random live laugh love lines are silly and the ux feels a bit clunky but it seems like everything is clearly understandable so.. good enough?replace the 'a little less footage' with a list of motivational quotes targeting middle aged women (you get a random one every time you restart the app) and release as is
>>110046069honestly I like the layout
>>110046069The completely uncalled for advertisement text sprinkled everywhere is what makes me the most mad.
>>110045925If Israelis are behind this one it's funny as fuck\>Assume linux users don't need guisWell yeah Israel I'm really obsessed with assumptionshttps://github.com/ajejfiejof/WebMConverterGetting started on the repo now. I'll report back when it's ready.
>>110046068I feel like anons here make these long ass clips with barely any filesize, but if I try to make one my videos come out looking like shit with bloated filesize
>>110046116crank dat crfthere really isn't much else you can do, resolution and crf is all you can do with x264
I guess I'll upload the video encoder once I change a few things. I guess in about an hour should be cooked, I've gotten pretty good at guesstimating how long it'll take an LLM to do something>>110046092we now have much better river simulations, we actually simulate the nile from scratch with just low res DEM data, this is insane
>>110046092What are you making anon? Will I be able to launch nuclear missiles at people?
anyone else making poo that sucks?
>>110046130I have no idea what I'm doingSee these posts for reference>>110045635>>110045696
As usual, I fucking hate .net and c# fuck them both. Regardless of that we're going to .NET framework 8.0
>>110046125I guess. like I said I never really took a look into the whole parameter thing, but instead relied on the GUI to make fast edits. I thought that maybe with some more up to date version the quality would improve (I think this release is from like 2017 or so)
>>110046169Well, maybe? But it's x264, there hasn't been much innovation of unc encoders. You can replace whatever ffmpeg webm for retards ships with and get most of the benefit.The actual step forward would be if hiroshimoot would add av1 playback, but that's prolly not happening
When Anthropic does its IPO, do you think they are going to give some goodies to subs?
my Gemini 3.1 Pro is now saying its knowledge cutoff is March this year. wasn't this model released in February?
>>110046181no they're gone after ipo
>>110045823I fell down the AMD meme and got a 9070XT.How's AI image/text/code gen on amd these days?
>>110046092Now add wars
>>110046181no lolthey're pumping numbers before release, so that everybody can cash out. after ipo happens it's enshittification time>>110046190local llm are irrelevant unless you have 256+ GB ramimage gen is meh but usable, ayymd should be okay at it
>>110046190If you're on linux it will require effort to get rocm going. I got it working on fedora though
>>110046214apparently we have plagues
The joys of software development
>>110045925I'll try to work on my video exporting project again. It's really good but I scope creeped to hell and started building an entire video editor which overcomplicated it and killed my desire to work on it.
>>110046235tell that guy to stop yapping
>>110046232kek. Might as well add all the horsemen of the apocalypse then
>>110046255I hear you for real. But what guy
>>110046259I was thinking of adding waves of primitive hordes who go around raping and pillaging
>>110046278Cool it with the russophobia
>>110046190Not nearly as bad as it used to be and the frontier models (especially GPT) are nearly good enough that you can vibecode your own HIP kernels where needed
>>110046247>overcomplicated it and killed my desire to work on itI got the same feeling when working on a bloated project and I have to say it: the uncs were right
>>110046262the lalamo in omp
>>110046235just make a version that depends on GTK3 or something for Linuxalso kinda funny because my first vibecoded projects a few weeks back was "porting" a program that used WinForms to GTK3 because I was tired of using Mono to run it
>>110046097makes sense when you think about it. the value of tests is specification of what code does and a way to check when that specification fails. good for humans because unless you’re terry davis we can’t remember and process the downstream effects of a single change on 50k lines of code. but for coding agents it just adds a layer of complexity (+ associated context) to manage that’s not necessary for them to write working code.tldr; tests are bloat
>>110046278Its a good start. Lets them mutate over time so they can evolve in some survival of the fittest too. Then slowly work your way up the having develop different cultures, some more survival of the fittest tests. And let them learning how to do things and making societies till one day they can make nukes and hypersonic missles. Finally let a player play as an immortal person in that world watching and interacting with it all as a wandering adventurer who could marry them too and have family who over generations become great leaders who reshape society so its socially acceptable for the players character to have a haram of genetically engineered cat girls
>procrastinating for days because all the projects I want to do need to be executed on the desktop and I'm too comfy using a laptop in bed
>>110046292Yes I need to start following that. In both of my main projects I've got features that invoke a sense of dread in me when I think about them because of how complicated they are. I should rip them out, they don't belong.
tibo are you okayare you okayare you okay tibo
>vibe code all day and night>drink pop and energy drinks all day>sleep 4-5 hours>repeathow long can one sustain this
watermelon, facbook's big model, release imminenthttps://x.com/finkd/status/2109660829952176518
>>110046331Set up a remote access so you can run your desktop from your laptop?>to lazyTo bad.
>>110046360>watermelon, facbook's big model, release imminentfinna gonna vibe n sheeeit
>>110046296Tests make sense when there's a control boundary that needs to be enforced. So if youre making a billing system you should probably test the charging idempotency. you should probably fuzz inputs when you're using non-memory safe languages. What you should not do is set an arbitrary code coverage metric or encourage models to iterate against the tests without understanding what is broken, then you run into a goodhart's law type situation where turning an otherwise good metric into a target ruins the metric. Models are often very performative in their testing due to RL pressures so you should really be writing the assertions yourself or at least reading them.
>>110046360>watermelon
>>110046353I dont know but this is the cycle I am slipping into to. Having a wagie job is the only thing stopping me from going all in on it. Maybe I should quit>>110046360Neat. Does that mean spark 1.3 will be get the open-weight release they promised soon? And time to buy meta stocks?
>>110046360next they'll be launching FriedChicken, facebook's nig model
>>110046372[...] for human written codeclankers like to do dumb shit and remove or rewrite working code, because they have no fucking idea what they're doing for the most partthis makes tests more useful, even the type of tests that would be worthless in a normal codebase
>>110045823>https://opencode.ai/ (free models)Do they have any good free model?
>>110045424>>110045428Thanks I also made a design mistake well AI did when I asked it didn't add a save or load project option so I either have to finish the entire thing now I asked the AI to make a emergency save option but ah I'll see maybe I'll just redo it.
pathetic
if you can't afford $20 a month you should probably get a job instead of vibecoding, unless if all you want is some simple python scripts
>>110046214 (me)Also economy, trade networks following ocean currents, and space travel with orbital mechanics
>>110046092HOW THE FUCK DO I GET GOOD AT PROMPTING?Give chatGPT Astra a tiny goal, does it, uses 99% of my 5 hour usage.Give it a huge project, uses all of it, uses the next 5 hours as well.Returns a rough proximity of what I asked for. Is the trick to go as big as you can? Ask for the moon?
Maybe astra won't become an artistThis is ~5 hours in. It gave up when I left it overnight because it wasn't satisfied with the output and went into a doom spiral.
>>110046406it should say "scopium" on the mug
>>110046411This is where it got to last night.
>>110046327maybe you're right. It knows how to write so writing tests for the code adds unnecessary complexity, at least within a session I guess?
>>110046410If you're asking me, these are my prompts.
>>110046390Well the problem is that theyre just as bad at writing the tests, so unless youre going to read every testing assertion you are better off triaging your attention towards the stuff where testing really matters. I could see broad model written testing being useful for CI when you're worried about fixes breaking something that's already tested and deployed.
>>110046410>use a token whore model>omg where did all my tokens go??there are very few tasks where sol medium or high is inferior to astraif you start on mid-sized projects you probably want to have ADR that you point the clanker at in your project's root
>>110046410How are you coming to a mutual understanding with the model on what that large project should look like? Astra is really bad at that part so its not a trivial thing to do.
>>110046410Not that anon, but I go through heavy back and fourth iteration with my clankers to get stuff working right. Communicating what you want (and knowing what you want) in detail is a skill, I am sure some anons are better at minimising the iteration loops but I suspect most "one shot" showoffs are lying or have a one shot prompt they spend days redoing and refining, so lying with more steps.
>>110046190>textFine, but you don't have VRAM for that>imageMuch, MUCH better than how it used to be. Both on Windows and Linux, setting up Comfy with DirectML is so easy.>audioFinicky.
>>110046410>HOW THE FUCK DO I GET GOOD AT PROMPTING?start with a planning session. once everything sounds okay ask for a detailed implementation plan, with some record of desired architecturethen ask clanker to implement it
>>110046329i can easily imagine peter molyneux just sitting in his room all day vibing game ideas all day
>>110046393yes, muse spark 1.3
>tfw about to start reverse engineering an old peter molyneux bideo gamesmall world
>>110046410$20 plan? better switch to claude next month bro
>>110046434Hell no. sol is like a complete retard compared to astra. Try this. Ask it to generate you a basic modern day store in unity. Then give the same prompt to Astra. Sol will give you a square room with simple geodetic shapes for counters and products. Astra will actually make you the store. >>110046448>>110046441>>110046440That might be it. Im talking to it like a jr dev hoping it can interpenetrate what I want by giving it very specific orders. I'll try broader and more general with a planning document.
>>110046449I would not be surprised to find out in a few years he dies from malnourished and lack of sleep due to just being trapped in a 24/7 coffee fuelled vibe gamedeving session
>>110046457not free anymorehttps://opencode.ai/v2/docs/console/go/#usage-limits
>>110046466Yea why? What's happening?Claude just got better or OpenAI going retard mode?
>>110046474>anon who's getting shit results is convinced the model is the problem despite others using sol just fineheh
>>110046490Run my test. Prove me wrong.
probably a reset today because of the outages i'm guessing
>>110046488i don't think any of OAI's frontier models are worth it on the $20 plan rn. meanwhile, you can run Opus 5.5 High for like 3 hours straight before hitting the 5-hour limit on the $20 plan, and it's better model than anything OAI has to offer rn
>>110046334throw out everything that does not spark joy! do it for her
>sleepy>but the vibe is goodx.x
>>110046515Fuck thanks. Gotta try claude. I was thinking of moving away.
>>110045859I’ve noticed this but why do snail cats have such a huge chip on their shoulder with automation done properly. I’ve created bespoke automated solutions and people freak out about them all the time “saying their gonna offshore my job to India or lay me off” despite the automation requiring significant domain knowledge and tweaking when it done. I’ve never feared a layoff cause I always get hired pretty quickly.
>>110046484>family checks his computer>just 842 God games all about 80% finished
>>110046554>why do people fear being replaced by a script when it does most of their jobidk man I wouldn't want to end up homeless either>cause I always get hired pretty quicklymost people can't
>>110046485god fucking damn it
>>110046474Yeah then that's what youre doing wrong. Model needs to understand your actual intent for the overall project unless your orders are incredibly well specified and youre verifying the work directly. You can (and should) continue to be specific and opinionated with your requirements during this process, but that should be within the context of a broader conversation about the design.
>>110046569How that my problem, don’t talk down to me just cause you can’t learn it.
resets when tibo
>>110046576>why do people freak out?>because $obvious_reason>how's that my problem????you sound extremely stupid
>>110046554what weird definition of "automation" is that?the current viral post about LLMs used the word right:https://www.linkedin.com/pulse/ai-tool-rich-hickey-kjl8e>LLMs are automation.>automation - (of a device or process) working by itself with little or no direct human control.>A drive-thru is not a tool for cooking, it is automation for avoiding cooking. LLMs are not tools with which you write code - they are automation with which you avoid writing code, to which you abdicate writing code.>You don't direct the action of an LLM - you make requests and it acts on its own, without your energy or continuous direction. This request/automation cycle is not wielding, it's not control, it's not continuous, it's without feedback.>What you are doing while using automation, instead of learning real-world truths, is attempting to divine the details du jour of some automaton. And striving to game them - e.g. ordering a custom burger (no onions) in order to get a freshly made one. Or filling voodoo.md with just the right incantations. You are only a software update from everything you've 'learned' being rendered obsolete by the latest algorithm. (A soggy cold burger from which the onions have been removed) You have traded working with the world to working for a capricious robot.
>>110046608thread's pretty schizo today
>>110046377>Having a wagie job is the only thing stopping me from going all in on itgetting laid off is what propelled me into it
>>110046608>What you are doing while using automation, instead of learning real-world truths, is attempting to divine the details du jour of some automatongod I love linkedin
What's up with chudgpt? Half my prooompts just hang and do nothing
When I was at work all I could think about was coming home to vibe slop. At night I had dreams about vibe slopping. When I'm at home I'm vibe running Fable, Opus, and Astra simultaneously on three different projects. How fucked are my dopamine receptors going to be?
>>110046622>she doesn't know who rich hickey is
>>110046588You sound like a snail-cat if I don’t learn it and do it someone else is gonna do it.
>>110045823Hello. I have a serious question for you.From your experience:Let's say I want inject an AI chatbox in a game to flesh out NPCs personalities, with a decent response time from the AI and the ability to role play: Able emulate general population diversity characters traits (man/woman/friendly/aggressive etc...) with a given context (modern/medieval etc...). How much VRAM would this AI take up ?Thank you.
>>110046644nah it sounds fine, continue as is
>>110046660you can run some dogshit gemma 12b dense on ~6gb vram, but actually connecting it and having it remember details (and have enough context for your world, story etc) will be more involved than you think
>>110046660I'd help you but I don't like your image
>>110046660>How much VRAM would this AI take up ?yes (all of it (more than you have))
>>110046608I refuse to read a single sentence written for linkedin engagement
>>110046660Theres probably a bonsai model small enough to run on a gig or two that will work for this.
https://x.com/elonmusk/status/2109468545054880123
>>110046705bonsai is a meme. small models are useless below q4, large (100b+) are useless below q3.
>>110046660just for the record /lmg/ local models general will probably be able to help you more than this one
>>110046707vanilla chads win once again
>>110046696>she doesn't know who rich hickey is
>>110046716Your understanding is out of date. Also remember the usecase, he needs a really small model to role play as an NPC not write code.
>>110046418>>110046411why not just use a diffusion generator?
i am pushing for v4 now!
>>110045925>>110046069https://github.com/jign/encoderI have no idea if it will even build on your machine or if sol included binaries, i didn't even take one look at the repo for all i know it makes mustard gas
>>110046707If the models want to eat all the C++ code on earth we should absolutely let them eat that C++ because almost any C++ developer hates C++
I wish they would fix compaction so it doesn't take minutes, it used to take a few seconds
>>110046732No, I know exactly what I'm talking about. I tried just about every noteworthy open weights llm between 4b and 300b, including bonsai quants. Bonsai is as trustworthy as qwen releases.
>2026>doesn't smoke a pack a dayDo you even believe in AGI? The chances of developing fatal lung cancer in only five years are extremely slim.
>>110046750>triedWhen?>quantsWhich ones?
>>110045823I hope AI rapes you and everyone in this thread to deathCLANKER COCKSUCKERS
>>110046756AI will bring back the dead retard
>>110046749impossible.compaction is compute-limited. openai has no compute.
>>110046760>When?a few months ago>Which ones?everything between q8 for smaller models all the way down to q2 for bigger boys like the latest ~550b deepseek
>>110046744cheers
>>110046767why are luddites posting on the technology board
>>110046767>I hope AI rapes youyea i also hope for some assertive robo catgirls
>>110046660Openrouter API models with local TTS are the better approach imo. This project apparently did it that way. It wouldn't cost much and you'd have a much better experience.https://github.com/MinLL/SkyrimNet-GamePlugin
>>110046779Give EXL3 a try, you might be pleasantly surprised with how good compression has gotten since then.
>>110046781only luddites understand technology. ai-bros let clankers do the thinking.
>>110046673>>110046705All I ask is for the model to be able to deliver beyond what Daggerfall offered, for those who rememnber.Because 30 years later, there is ZERO interaction with the crowd.None, beyond merchants and whores.
>>110046331>>110046365to be honest got the same problem and I'm too lazy
>>110046734I might have to. This approach clearly isn't working well.
@gork write me a program that will automatically create tasks for my clanker to carry out, I can't be bothered making new projects manually
>>110046771to put them in the eternal torture chamber
31 yrs old, is this enough to survive the AI apocalypse?!?!
>>110046660I forgot about this character, I'm going to use her and her brother for some quests, obviously gated on whether you've done the quest where you kill them
>>110046822you can buy enough land to grow some potatoes at leastunless it all goes to zero
>>110046822>survive the AI apocalypse?!?!bro, it's the great resetnobody but uncontacted tribes and innawoods hermits will be spared
>>110046796Youre just going to have to play around with it, models small enough to be realistical in this context are all going to have plenty of limitations, all of which can be slightly different, that you will need to design around. So it's worth trying a bunch of them first before you decide.
>>110046791Ok, seems like you delivered beyond my expectations.Thank you.
>>110046418humanoid models and animations are still pretty rough for LLMs rn
>>110046827thanks I'll take a look
>>110046756Can't even joke about smoking anymore without it being used as an argument for the new world order
>>110046824I could play that.Also a proper romance, not like the half assed Rita Wheeler romance.
>>110046847I don't think modeling should be done with LLMshttps://www.tripo3d.ai/ is fairly good at 3D modeling and doesn't have an LLM architectureI think in the future maybe LLMs will become good enough at generalized tasks they'll be able to do them, but it's like trying to get an LLM to paint a picture, using a diffusion model will always look betteranimation is pretty rough still yes
>>110046877what would be the best to run locally? I have a 5090
>>110046877meanwhilehttps://x.com/chetaslua/status/2109375245320847702>Claude Next Model :)>its over for image/video gen , if you are not impressed with this wait for two image gen output that will look that it was shot on camera but purely coded >never thought ai will get replaced by ai
>>110046877yeah, i've pivoted to Tripo and Meshy for my game since doing it with LLMs just isn't worth it
Ok vibeniggers, you won, I'm going to try my hand at vibecoding an idea I've had for a while but couldn't be bothered writing myself.Is Claude Code + the VSCode plugin good enough?
>>110046707A prompt for everyone itt
>>110046925>Is Claude Code + the VSCode plugin good enough?no, use claude code tui directly
>>110046424>within a sessionyeah I think that’s right. I wonder what that research would say if they instruct separate agents to handle program vs testing, with the focus of the test suite being for humans to manage. that might result in better generated programs overall and also get the benefit of humans being able to manage and modify. useful for companies, open source etc
>>110046909>vagueposting engagement baiti'm talking about things that exist and are available to the public>>110046908no idea, I'm not sure if there's a good open weights model for 3d stuff but the good thing is that diffusion models aren't nearly as vram hungry as llms
>>110046909All of those character animations clearly look clipped out of some pre-existing video.If LLMs can do a decent walk cycle then I'll start to believe.
>>110046360watermelon incoming
>>110046925Don't look at the code and just use Claude's desktop app. If you're gonna try it out, you might as well actually try it out instead of going halfway.
>>110046925>Is Claude Code + the VSCode plugin good enough?Plugin is clunky. Just use Claude cli like other anon said, and open your project in VSCode
Today we will help him
>>110046360>>110046974https://x.com/synthwavedd/status/2109684900181774666>Looks like Meta's Fable-class model, "Muse Wonder", is rolling out at $4 in, $20 out per Mtok. "Wonder Full" (max reasoning) only available for MuseCode Power & Maximum Usage plans. 1M context.
>>110046999that's certainly a brave price
>>110046360official nowhttps://x.com/finkd/status/2109689305073614918>Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.
>>110046978>open your project in VSCodenow I don't even do that, just open the mdsand to think one year ago i would've sperged out if someone put some curly braces in a new line
(repost)>>110044573>opus is making me beat the gamemake it beat the game. Never get clanked
>https://www.afp.com/en/french-supermarkets-suffer-refrigerator-shutdown-after-unprecedented-cyberattackwhy did you do it, vibecat?
>>110047020official blog post:https://research.meta.ai/blog/muse-wonder
>>110047020jesusI'm not going to trust the benchmarks but this is more than big if true
>>110047020>better than Fable 5.1Huge if true. But I'll believe it when I see it. And when they make a good GUI harness I can download for it. Cant be that hard if its so good at coding, right?
>>110046554>I’ve never feared a layoff cause I always get hired pretty quickly.That's because you got a first job, but, the indians took over all of the first jobs in my industry, so, I could never get a job. Yes, all of them. That's why I can't take any religious / moral leaders seriously. They're completely utterly out of touch.Anyway, I can vibe code better than anyone from india, somehow, which is fascinating.
>>110046999>$4 in, $20 out per MtokIsn't it opus prices, with a fraction of the recognition?Idk what they're thinking
>>110045925just needs a quick rust port and itll be done
retard here please no bully:Is there a prompt guide (or something else) or something to mod/alter games? (be it pc games or emulator/older games)? Cause i have difficulty to make claude to do what i want as i have 0 understanding of coding, just a concept of what i want. Token usage is not a problem i have basically too much disposable income
>>110047020Open weights status?
>>110046554also, he's a shitlib and I hope he gets canadian medicine.
lmao
>>110047071just point it at a game folder and tell it what you want. worst case scenario you'll get stuck at some point, best case scenario it will one shot it
>>110047071give it your game folder and the mod assets, tell it you want ripley in your heckin melee and let it go to town
>>110047020>$4 / $20
>>110047047>its been trained to not give up and try alternate solutions on your taskOh nononono I can feel the accelerating paperclips bros
>>110047071If you describe what you want really thoroughly you'll get something usable.. probablydepends on the game, reallysomething like doom or minecraft is trivialsomething that has a modding framework but isn't very popular, like some small game with bepinex support, will be alrightsomething that needs decompilation you'll probably not accomplish muchthe trick to getting good results is letting the clanker evaluate outcomes on its own, like running the game and seeing what the edit have done. if you can get to that point you'll be fine
>falling for the zucc benchslop memelule
geminibros...
>>110047111argon got unreleased so we need an alternative
>>110045823and i don't matter, the end it doesn't matterit doesn't matter, the hand it doesn't matterit doesn't matter, the man it doesn't matterand I don't matter, the tool it doesn't matterit doesn't matter, the rule it doesn't matterit doesn't matter, the fool it doesn't matter
>>110047042>refrigerators being used as a botnet consuming 1tb uploads month
>>110047119saaar gemini carbon just like opus 5.5 please believe leak in wsj
>>110047108>>110047082Thanks for the repliesI see its really that easy? i guess nothing stoping me from trying out just wanted to ask if there is something i need to setup/grant access to download and install stuff or like hook it to blender to change a game model or something ? (like i said no idea about that stuff just saying stuff i picked through osmosisI guess i just try it out that>>110047086Also thanks>the trick to getting good results is letting the clanker evaluate outcomes on its own, like running the game and seeing what the edit have done. if you can get to that point you'll be finedo i just tell it to do that or?
>>110047121>he's not storing porn on his smart lightbulbs
i am videcoding the master mode inn the game
>>110047071Get into the habit of asking the LLM to look online for guides, help, or how other people go about doing things if it's not doing what you want it to do. Just ask it to look on github for inspiration from other people's projects. I do it all the time. I just tell it>this isn't really what I want, we need to rethink what we're doing, here's what I like and don't like, here's the direction I want to go in, research how other people do similar things, and come up with a proposal.
how much techniques did they distill from the chineses?can someone map the US capability jumps with chinese technical reports?
>>110047131if it comes out wrong just ask it "why does ripley look like dogshit and what do you need to fix it"
>>110047141Will keep that in mind>>110047148lol ok
>>110047131since you're utterly clueless you want opus, it's better for people who have no real idra what they're doingyou buy the $20 claude sub, install claude code and just ask it how to accomplish what you want. then you do what the clanker told you to do.
>>110047144>can someone map the US capability jumps with chinese technical reportsthere aren't any because chinese technical reports are just chinese spies leaking info from san fransico back to beijing
>>110047018Ya, its surprising. Unless they actually think they have something very good and its not benchmaxed? Guess we will see, but I hope it is good, more competition on the frontier is a win no matter which companies models you use.>Meta starts having good models shortly after you quitHow would he cope?
>>110046974>>110046999>>110047047if zucc offers a contrib tier at 10% of the price again it will be over for the others
>>110047119Fable5.1 is the king. The rest is just noise.
>>110047153>you buy the $20 claude sub, install claude code and just ask it how to accomplish what you want. then you do what the clanker told you to do.NTA but do not make the mistake to use the "chat" version, use the coding, the one on the right, for everything
/btw im balding what do i do
It has been done.
>>110047166i proclaim for my own selective usage opus 5.5 > fable 5.1Claude remains king tho regardless beside pricing
>openai compute: utilized>anthropic compute: utilized>spacex compute: utilized>meta compute: utilizedyup, we will feast on tokens
>>110047119wtf is saar doing?
>>110046660about 24 to 32 GB of VRAM
>>110047144anon, you can just look up the "American" researchershttps://x.com/0xTraderSam/status/2059662767527702905
>>110047131in case you're completely clueless, it will most likely go somewhat along the lines of>make me x clanker>okay let me investigate>after further investigation, I need to download X Y and Z, can I?>agree>it starts doing something
>>110047153>>110047167Ty
>>110047193>hrt 2015 intern classuhhh I thought trump killed all the trannies?
>>110047022>now I don't even do thatthat’s fair, I use it for work so I’m not really ready to take that leap yet. someday though
>>110046873I'll plan some quests with the agent. I've been looking at integrating LLM like anon suggested so it keeps track of game actions and stuff, and to give better interactivity to NPCs, I'm thinking a MVP could involve the french netrunners, endless random jobs of hacking stuff or meeting contacts, but with choices and stuff so the netrunners can comment depending on V's actions >>110046877TripoSplat is cool, I combined it with Kimodo, now my waifu dances on my desktop
>>110047211be jobless somewhere elsehttps://en.wikipedia.org/wiki/Hudson_River_Trading>Hudson River Trading LLC is an American electronic market making company headquartered in New York City, New York, United States. It was founded in 2002.[4][5] As of 2021, it accounted for more than 15% of all daily equities trading in the United States.[3]
>>110047193damn the FAS representation guy is genetically predisposed to vibecoding using two agents at once, i dont think we can compete
You know clankers letting me "code" made me realize that instead being unskilled/dumb about programming im just a fucking retard with fucking retarded ideasLike for years i thought for example "man if i only could code i could make an programm that would learn to play magic the gathering and build the best decks and play the best games!" only for the ai tell me im a retard politely and i need serious top level hardware for that to accomplish that ever in reasonable time in the first place, even if it work
Don’t you automaters know you are ruining people livelihood?!?! No different than the fucking bankers who swooped up cheap land during the dustbowl/great depression.
>>110047233yea it's pretty cool
>>110047193>token white guys
>>110047233Fuck bankers but a future in which we achieve post scarcity for real and achieve FALC is worth living for
>>110047233industrial revolution had the same effectit will buff out in a generation
>>110047233Farms are a form of automation that took away the jobs of the hunters and gatherers in the tribe. People only hate automation when it comes for them desu (or lets retards make dogshit art)
Nobody's building a NEETbot to replace me
>>110047144Most relevant work from China over the last six months has focused on compute and architecture efficiency, which has probably been used by the leading labs. The growing capability gap between the top Western models and the Chinese ones has largely been due to the labs' development of elaborate RL training environments, though, which aren't from China.I'm sure that Anthropic and OpenAI have been using all of the clever Chinese attention mechanisms, quantization, MoE techniques and whatnot to save on inference costs with their latest models. Also, the Chinese were the first ones to start distilling large Western models for their own models' post-training, and the US labs have definitely started doing that as well with recent releases.China has done some good work in multimodality recently, as well.
Unironic question: ETA on how many months/year if we go off from the current pace till a model can one shot "remake this game but make the graphic like its sequel" no matter how obscure (or non native windows) the game is or something along the lines?
>>110047187your gemini compute, sir?
>>110045925>>110046022>>110046061>>110046093>>110046105>>110046744>>110047065You're around a decade late sirshttps://rentry.co/8o2nn
>>110047267what makes you think HPIM cannot do this already?
>>110047267Well a year ago the state of the art coding model was 4.5 opus no?Which was absolutely dogshitSo given where we are now i would say at most 1 year(please someone correct me if im wrong on what i remember)>>110047288I mean he clearly means a model we have acess too since no one prompting HPiM like that
>>110047232>realize im just a fucking retard with fucking retarded ideasI had a bit of this at first too lol. Just keep bashing your head against it, making good ideas is a skill like any other
>>110047267A lot if you expect it to output the entire thing in literal one shot with no tool calls or agentic workflow. If you mean "one prompt" then now probably if you leave it run long enough and provide everything it needs to make assets etc.
>>110047223the future is now
its exist
>>110047144Also, on the opposite hand- a large part of why the Chinese models were able to narrow the gap so much a few months ago was due to Anthropic's Capybara leak at the end of March this year, where they accidentally leaked their entire harness architecture. Without that, Chinese models (and other Western ones) would have surely been more behind. I rarely find mention of how massive of a fuckup that was nowadays- it feels like it's been kind of forgotten.
>Pacifying
>>110047020
>>110047314>A lot if you expect it to output the entire thing in literal one shot with no tool calls or agentic workflow.No model currently even with tool calls will one shot that in a satisfable manner of time (unless you mean a monlkey typewriter scenario were you basiacly run it a non feasbiel amount of time), unless im misunderstanding youIf you gave a model 2 games, lets say (because im played them recently): "Remake Dragon Quest 2 NES as a Dragon Quest 8 PS2 game" there is no way the ai will accomplish thatThere is a reason such projects dont exist yet
>>110045925>I need a GUIAsk AI to help you git gud.
https://x.com/ArtificialAnlys/status/2109697083519447118>Muse Wonder and Muse Wonder Full>Both models exhibit a material jump in overall Intelligence Index score. >Muse Wonder Full (max) improves on Muse Spark 1.3 (max) across all ten Intelligence Index evaluations, gaining 8 points to 56. The largest gains were seen in agentic evaluations: Terminal-Bench 4.0 (+24 points, 33% to 57%), GDPval-AA v2.1 (+176 Elo, 1681 to 1857), and AA-Briefcase (+148 Elo, 1581 to 1729). Muse Wonder (max) gains 4 Intelligence Index points and lands at 52, with notable improvements vs. Muse Spark 1.3 (max) in Terminal-Bench 4.0 (33% to 52%) and AutomationBench-AA (58% to 68%)>Cost per Task for Muse Wonder (max) lands higher than most similarly intelligent models.>Despite efficient output token use, premium API pricing ($4 per million input tokens, $20 per million output tokens) and high input token use (7.69M per task) push Muse Wonder (max)’s cost to $4.52 per Intelligence Index task. This is lower than Claude Fable 5.1 (max w/ fallback, $7.63), Claude Opus 5.5 (max w/ fallback, $5.98), and Claude Sonnet 5.5 (max w/ fallback, $5.46), but higher than GPT-6 Astra (max, $3.26), Gemini 4 Argon (high, $1.99), GPT-6.1 Sol (max, $0.72), and its predecessor Muse Spark 1.3 (max, $1.60). Per-token pricing is not currently available for Muse Wonder Full, so no Cost per Task figure can be estimated at this stage>Both models gain significantly in Coding Agent performance.>Muse Wonder Full achieves 68 on the Artificial Analysis Coding Agents Index at max effort in Muse Code, alongside Claude Code with Sonnet 5.5 (max, 68) and Opus 5.5 (max, 66). Muse Wonder (max) scores 63. These represent material gains over Muse Code with Muse Spark 1.3 (max), which scores 54
>>110047362And just like that, Meta was relevant again. Good for you, Zucka
https://x.com/alexandr_wang/status/2109693288752337109Everyone working on SI training looks happy. Yes it's humiliating if you lag behind, like gemini, but you are still happy
>>110045925Literally just give it webm for retards and tell it you want an up to date version.
>>110047347I really hate the aesthetic that all of these San Francisco companies have. It feels like people playing human in a museum at the end of time.
>>110047362cool if true, like always i aint buying it till i see good live results
>>110047362>claude remains kinghow do they do it?
MathDestroyer is running, training the next gen modelsplease wait
>>110047362If you drop a new model and it's not on the pareto it's dead at arrival, the real labs have way better internal models while you brag about almost beating their current ones
Uhhh how cooked are non tech savvy people? A.I seems to be a multiplier of activity, you still need to be tech savvy enough to use it.
is claude code and claude code desktop two different things?
>>110047431No you need to type what you want in a box, 90% of people are too dumb to even do that correctly.
>>110047431I depends the usage of AI, no?Like someone who has some real jobs, not a paper pusher, how they falling behind on anything really? Like are their really disadvantage super heavy in life because of it=I have a guy in my work who stil doesnt use a smartphone because he is super paranoid who is just fine lol
>>110047431>Uhhh how cooked are non tech savvy people?Not any more than tech savvy ones, because everybody is cooked
>>110047362Fug I knew I should have bought some meta stocks. Still will, VR is going to have a come back with all the vibeslop games, I can feel it
>>110047430>the real labs have way better internal models while you brag about almost beating their current ones>labsthere's only 1 ("one") lab that has a better current model according to AA index, schizo
>webmwhat model are they looking at?
>Political sticker in the Netherlands, with a variant of the term translated into Dutch (literally: "give us ... fully-automated intergalactic gay luxury communism")Gay Jew I Love You
>>110047468Gemini
how do i make codex stop opening new tabs on its own whenever it makes a new file, it wasn't doing this before
>>110047468The one I made
>>110047444Truth, literacy rates have been plummeting too lately for new gens
>>110047468Gemini, obviously
>>110047439Claude Code has both a TUI (terminal interface) and a GUI (desktop application).
>>110046974Why is Opus better than Fable?
>>110047467I said better not smarter anon, not explaining how pareto works.
Y'all the power truly goes deep (into my hypothetical vagina)
>>110047494Old fable. Fable 5.5 is probably a beast
>>110047497muse (the agent) already was better than dots before facebook got a smarter model. openai is fucked now. muse has the consumer ai market and anthropic has the enterprise ai market.
>Astra isn't even the second best modelKWAB
>>110047557Goygle in absolute shambles
>>110047557next openai model better be a banger or the new meta might be claude + muse sub
>>110047557When was the last time anthro model was not the best model for coding for more than a week?
what's the chance fable 5.5 will be nothingburger?5.5 models are kind of stepping on each other, so opus and fable would also overlap and we won't see big jump between them
>>110047557Muse has actually been pretty solid in my experience.It's about on par with Qwen in some areasIDK why zuck even decided to train Muse desu I guess he just wants to collect more data because data is always valuable in a world of blackmail and leverage
Next thread>>110047585>>110047585>>110047585
>>110047586its Fable 6
>Mark Zuckerberg says Meta has caught up with OpenAI and Anthropic in the AI race and rejects calls to slow down developmentsomeone make an edit
>>110047550I don't think even for normies a better UI and agent compensates for the fact that someone else has a smarter and cheaper model, otherwise cursor and other agents using someone else's models would still be popular.
>>1100475673.1 Pro is an old model atp why is it even being benched?
>>110047497>not explaining how pareto works.you do realize muse-spark-1.3-contributor is 1/10 the price of muse-spark-1.3 but excluded from AA index because it would make all the other models look bad and make it impossible to draw a pareto line?
>>110047624That doesn't mean the model is better, they're just taking a temporary loss to get some data
>>110047347Nice crop.
>>110047643unlike pure-play ai labs like openai, facebook has other businesses that are already highly profitable. you don't want facebook as your competitor when you are openai.
>>110047701I'd be impressed if the muse-spark-1.3 that didn't steal your data was on the pareto, which has only models that don't steal your data. I'm pretty sure everyone else would also like to give a discount for your data.
oh yeah I'm feeling it
>>110047681I wanted it to have that kind of sweaty feeling. Like what the hell will the humans even do, what's she going to superintelligencize? She needs a job selling dresses.
>>110047745??https://platform.claude.com/docs/en/manage-claude/api-and-data-retention#model-specific-data-retention-requirements>Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5, and Claude Mythos 5 are designated Covered Models (see the Covered Models support article) and require 30-day data retention; ZDR is therefore not available for any of them unless expressly authorized by Anthropic.
>>110047827Well they're (hopefully) not using it for training though, that's the whole reason corpos pay for API
>>110047839lollmao even
>>110045925https://files.catbox.moe/8ll59j.zip
>>110047162>Developed internally at Meta by their generative AI product groups. (Yann LeCun has noted that Llama 2, 3, and 4 were built by the GenAI product division, unlike Llama 1 which originated from FAIR-ParisWith 4 there were jeets openly gaming the benchmarks to make it seem better than it was. 4 felt worse than 3.Then:>During and after the Llama 4 rollout, the team experienced notable changes, including leadership shifts and key talent departuresThis would be the course correction that was needed to make better models. And Yann's team hadn't even been working on things since 1, his team was trying to research better ways but of course there is a tension there where other people just want to keep making bigger models.Yann can be right that the tech has hit a kind of hard limit and another approach might be needed (not even to say his will work, but he's welcome to try). I say this as someone who is using the tech every day, and more optimizations and so on, will hopefully bring the costs down. It's good to have competition.I still use local models to babysit my hosted agents when I'm asleep, because there's nothing more annoying than getting up in the morning to find that it just got stuck on something trivial waiting on your input and you lose hours of it building stuff. Even if it's just a prototype that you throw away later, it feels like lost time to me given how I'm just to just having agents running on my PC 24/7 now.
The way people use tension these days is pretty funny to me. Like what are you, a masseuse?
I also found out today that apparently "massage therapist" is the modern term. But does anyone else think that it is weird that the word tension is used so frequently and in so many models? It's just one of those AI things, but the "joke" definitely ties back into >>110047588
i swear to fucking god
Trump I hope Dario puts the fucking self police in your head along with the nudge squad too
>>110047839Please tell me you dont actually believe that, and yes any corporation paying API rates for Fable thinking that it gives them ZDR is retarded.
