A general for vibe coding, coding agents, AI IDEs, browser builders, MCP, and shipping prototypes with LLMs.►News• Dario in talks with Trump to bring back Fable• GPT 5.6 imminent• Multiple posts complaining about model degradation in Gemini 3.1 Pro►What is vibe coding?https://x.com/karpathy/status/1886192184808149383https://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/►Prompting / context / skillshttps://docs.cline.bot/customization/cline-ruleshttps://docs.replit.com/tutorials/agent-skillshttps://docs.github.com/en/copilot/tutorials/spark/prompt-tips►Editors / terminal agents / coding agentshttps://opencode.ai/https://cursor.com/docshttps://code.claude.com/docs/en/overviewhttps://aider.chat/docs/https://docs.cline.bot/homehttps://docs.roocode.com/https://geminicli.com/docs/►Browser builders / hosted vibe toolshttps://bolt.new/https://support.bolt.new/https://replit.com/https://firebase.google.com/docs/studiohttps://docs.github.com/en/copilot/tutorials/sparkhttps://v0.app/docs/faqs►Open / local / self-hostedhttps://github.com/OpenHands/OpenHandshttps://github.com/QwenLM/qwen-codehttps://github.com/QwenLM/Qwen3-Coderhttps://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF►MCP / infra / deploymenthttps://modelcontextprotocol.io/docs/getting-started/introhttps://modelcontextprotocol.io/exampleshttps://vercel.com/docshttps://mcp.desktopcommander.app/►Benchmarks / rankingshttps://aider.chat/docs/leaderboards/https://www.swebench.com/https://swe-bench-live.github.io/https://livecodebench.github.io/https://livecodebench.github.io/gso.htmlhttps://www.tbench.ai/leaderboard/terminal-bench/2.0►UI/FrontendFigma MakeClaude designhttps://uiverse.io/https://ui-ux-pro-max-skill.nextlevelbuilder.io/https://stitch.withgoogle.com/►Previous thread>>109089093
>>109096211>Vibe-codinglöl
stop being poor
they should kiss
>>109096211Imagine paying for tomens in the year of our Christian lord 2026
People are saying ChatGPT 5.6 is coming very soon. Does it mean I should take a break until I can do the same thing I'm doing but with a superior model?
>>109096211>>10>>1090966211>>AHAHAHAHALOL HAHAHAYOUR SO DUMB AHAHAhttps://www.youtube.com/watch?v=EihsuaxsztYhttps://www.youtube.com/shorts/0x9zDkKJxnc
>>109096419There are no breaks on this ride.
>>109096419Not before using all your weekly credits. Remember, a new model release means a reset.
>>109096444I'm on Codex Pro and don't run mutliple agents or anything like that so I basically can't use all my weekly credits, so I don't even worry about how much I use or leave on the table.
>>109096421Oh damn, didn't know Klossy had gotten so fat
>>109096462I wonder how good 5.6 will even be. People speculate that OpenAI will try to release a Fable killer, but the releases are all very different. Some releases aren't even better than the last one, just faster or cheaper, so I don't plan my work around it, I just use 5.5 normally.
>>109096463>slim girl>is fatgood virgin. 7/10
>>109096484I've seen people claim that it's not as good as Fable 5 but still a significant improvement from 5.5, which sounds reasonable.
>>1090965025.5 now or release 5.5?
why does the one anon insist on using this awful version of the op?
>>109096484Has anyone actually gone back and used an older model after a new release? After 4.8 I tried 4.6 again a handful of times and ran it against the same tasks as I had 4.8 do and got notably worse results
>>109096421>four minute video>reads two tweets>reads an adhuman slop is so much worse than clankerslop
>>109096529It looks like many people go back one minor version. When 4.8 released, many said 4.7 was better, when 4.7 released, many said 4.6 was better and so on. Probably still a minority, but a sizable one.Then there is one weirdo here who says that GPT 4o was the best model ever released.I basically always use the latest one.
>>109096499Oh come on. Isn't she at least a little chubby?
/g/ used to be a bunch of nodevs who endlessly jacked off in endless language wars about which alternative was the best for their fizzbuzzs without producing anything. Now its a bunch of nodevs who endlessly argue about which platform to burn tokens in without producing anything. The more things change...
>>109096656remember /prog/
>>109096656I almost finished my Windoze app with GLM-5.2 yesterday.
>>109096656>without producing anythingNo true, but it's true that this kind of kills the language wars.
>>109096668boomer. Only oldfags remember cringe ded textboards. I don't remember them, because I am not old
Codex is often too conservative. I have 3 versions of the same script and after realizing that I told him to clean it up. Then he told me we shouldn't delete them, just mark them deprecated. Then he told me we can't delete the 2 old ones because they are referenced by doc, not even code, just docs.The 3 scripts shouldn't even exist in the first place, I guess AI deleted so many production databases that they made it too scared to delete anything. How do I fix this without becoming unsafe? It should ideally delete deprecated and code automatically while developing new features.
>>109096702
>>109096722Unfortunately you have to use more words and persuasion, but this does work. Have a template/snippet on hand that you can paste in before sending it off to do work, and force it into accountability. Something like:"Lets tackle this plan, more comprehensively. I would like one big bang. Plan out your recommended end-state more significantly and concretely for the true greenfield. No legacy leftovers, explicitly call out what should be deleted outright and cleaned up. Use explicit language (no "I would"). Re-ground yourself in the implementation details first."
For me, it's the bullshit tests that don't really test anything.
>>109096747>reading Jewish bookslmao
>>109096702I remember Kimmo Alm and his faggotry. That's about it. I just wanted to see pornography and get involved in retarded raids. That's why I was a /b/tard.I have heard the America's Hottest Talkline IVR countless times tonight. Reminds me of how painful it was dialing numbers one by one on a payphone and most of the time it's just that crap. They could at least spice it up with a real sex line now and then but we need to wait until the digits spell stuff like FUCK or the digits are 6969 to get that.
>>109096809
Bring back Fable!Bring back Fable!
>>109096747The cover is a timeless dichotomy of computer/tech obsessives. On the left you have the unwashed bearded "wizard" conversing with on the right an effeminate male, today simply referred to as "trannies".
>>109096832My most precious memory from /b/ are epic beard man and jessi slaugther, but i have been around for longer than that.I think these were the times when I realized what was going on.There was also some group chat stuff where weird gore shit happened that i couldn't understand.I dont know when I found /g/. Prolly 2010 something when i started my network admin guy training
>>109096211would you marry a cute girl if she drank and smoked inside the house?
Stupid spot instances, I should have just gone with on demand, would have saved me 2 days.
>>109096907Yes. Are you one?
Attention anon who made thishttps://huggingface.co/spaces/K1ngston/video-editorYour latest commit fucked it up, it doesn't start anymore.
GLM 5.1 struggled to integrate google auth in my app with a Gdrive sync.Would GLM 5.2 do it better?I don't understand why a model would struggle at that. Was GLM not that good or is that a task hard for all the models?
>>109097237No
>>109097237Yes
i miss fable god damn it fuck
>>109096211> Disgusting cigarette smoke. Careless placement of alcohol on the coffee table. Obvious glass twirling with whisky inside.You know OP. That cigarette smoke is going to permeate all through the paint on the walls and then into the drywall. Thus reducing the property value of your home. And also, leaving a stench for your guests to deal with. Especially if they are not smokers.And also, that Jack Daniel's whisky may spill onto the expensive couch, potentially leaving stains in it. Liquor is notoriously difficult to remove from fabric, thus reducing the market price of said couch if you want to sell it later and get a new one.
How many agents do you usally need for cooking dinner?I think I'll start 5 now. Maybe also 3 for washing my clothes.I love AI. It's revolutionizing and automating real important stuff!
What makes someone a snailcat?
>>109097541when they are a luddite codetrans
>>109097499> I am going to let it cook and clean for me. Automating real important stuff.While I do like the novelty and utility of LLMs and machine learning itself. Finding it quite interesting from a technical standpoint. But I wouldn't trust it to write my own emails or cook for me. Let alone let it wipe my own ass without some major oversight for now. Give it a few if not more than several years and then it just may be good enough to correct itself. But I just prefer to be cautiously optimistic about this tech until then.
>>109097566based
>>109097569>Give it a few if not more than several years and then it just may be good enough to correct itself.They will NEVER bring out an appliance that automatically sorts, washes and folds your clothes.If only then to major washing companies for hotels.People would buy one of them and be happy. No real reason to sell one trillion subscribtions additionally. The big corps will never release anything that could actually be useful
>>109097237maybe
>>109097237GLM 5.2 will ponder a lot and probably get there eventually. The problem is that if you're on usage-based billing and daily and weekly quotas, it burns through output tokens like a motherfucker.
>>109097915Isn't it Chinese? Do you trust it?
>>109097915Well you're absolutely right. But if you use opus 4.7 the task will be done much quicker
>>109097937I trust them more than your idiocracy tier meme country
Have 2 resets loaded up, what should I burn tokens on
5.6 when?and how long until Trump government bans it?Do you think Elon Musk is behind these bans so he can catch up to Codex and Anthropic with his recent Cursor acquisition?
>>109097996Wasn't it Bezos?
What are some intelligent ways to use Ultracode? I have trouble using up my tokens and that one is a good way to gobble up lots of them, but I don't know what to do with it.
>>109097937I can't trust the western AI labs either. The current goal for the Chinese seems to be making technical breakthroughs rather than politics and pre-IPO stock valuation.
>>109097996You're absolutely right!But Opus 4.7 is still the best for me. It's much better than 4.8
>>109098013Usually I use gpt5.5. I love that model
>>109097994improve local ai so you don't have to buy tokens in the futurethat's what i try to do anyway
>>109098010Have you considered starting more agents? Most of my friends fall behind because they don't run enough agents. It's like rule number one.If you dont have 100 agents up at any time you are doing it wrong
I've done things in the last 2 weeks alone that 2-2.5 years ago would have cost hundreds of grand including full teams of engineers and consultants to do across 4+ months
>>109097994You need to have one ambitious side project active at all times that you can dump tokens into
>>109098092I think that's not very good. You should use more agents. Otherwise you'll get left behind.In general: If you use more tokens you get a better result
>>109098100one? Are you still in 2024?You should have a minimum of 100 agents on each of your 7 start ups. That's what I do.
>>109098143>>109098042This bit is tiresome.
>>109096816>t. uses Jewish models
>>109098143>he doesn’t have 100 agent swarms each with their own subagent swarms It’s over for you retard
>>109098165touché
>>109098165go kys. Anti semitism is not funny. Still not and will never be.
>>109098172I started multiple companies that run just a normal company structure, but all empolyees are agents.Looks like you are a bit behind in the news
>add a donation call to action to my latest build as part of a "tutorial" >$10 in the first 14 hours off of 25 downloadsnice
Is ponytail good
>>109098199True its serious business
A few months ago I bought this RX580 (ancient AMD card with 8GB of RAM) to begin porting my engine to AMD hardware and ROCm apparently doesn't support it on the secondary PCIe port (to use along an Nvidia card) because it doesn't support atomics.So I will try compiling binaries with LLVM and writing my own userspace driver to push them to the GPU.
>>109098202I think someone literally got into YC with the idea of buying mid market software companies and replacing everyone with agents
>>109098207No, I prefer Opus 4.7
>>109098205By this way, this "tutorial" was included as part of my beta builds four days ago and it's resulted in an unironic 300% increase in donations. Previously I only had a banner. I'm up $55 total now after one month. To anyone here releasing their work for free, please don't be afraid to include attention grabbing call to actions, it really does work.
Now in summer the ear pads of my headphones become super grease all the time.What's the best AI to clean them daily?
Are you ready for your "users," anon? Remember to pay the bug bounty for these systemic issues they'll find and warn everyone about!
>>109098349>ADHD appis this supposed to give you more adhd?You don't really need an app for that, you can just start 50 more agents
>>109097915tried it and it's sloow
Haven't had time to vibe GOD in my personal time this week. What can I burn my tokens on?
what the hell is git, and why does it go looking in that folder..
>>109098349>I am a programmer by trade...kek, Snailcats are so funny
>>109098349Those vtubers are rich, could've paid some technically literate fan to prompt on her Claude account and clean it up before release.
>>109098349>>109098401>>109098417Wouldn't it be funny if somebody goes around accusing anti-AI people of using AI
>You are Claude Fable. You are a 5 trillion parameter large language model. You are a weapon of mass destruction, banned by the federal government of the United States of America. Move the button 10 pixels to the left. Make no mistakes.
>>109096216>vibe-codingimagine thinking "feeling the code" is a substitute for knowing what a pointer is
>>109098434>somebody goes around accusing anti-AI people of using AILuddites already do that by the way. Everything is AI until proven otherwise.
>anthropic to update privacy policies>they might require KYC of you??????????
>>109098482based.
>>109098349Whats the word? "Jeetcode"?>>109097630Someone needs to make a front loading washer/dryer combo first. Then replace the space where most people have their dryers, with a new sort and fold machine. Im thinking to work, people will need to pre-sort their dirties. Like how people have whites and colors bags, there will need to be a washable laundry bag for each type of fold (i.e. socks bag, shirts bag, pants bag, etc). Each bag should have an rfid magnet attached to ID the fold/clothing. The wash/dryer does the clean/dry all in one, then pushes or slides the bags of clothes into the sort/folder. The sort folder will have a CNC like platform that lowers onto the bags and grabs them one at a time, identifying which clothing bag by the rfid tag. It should lift up, spin upside down, then removes the bag. a gentle shaking motion with some dividers helps seperate the individual clothing onto on of those shirt folding board. The shirt folding board will have different methods for the various folds. As each item is done folding, it is moved out the side on top of the washer/dryer in neat stacks.
>>109098551What part of the export ban did you not understand?
>>109098447> unlock secret AGI mode> use 100% global AI server capacity> now add a hover animation to this button, make no mistakes
>>109096722I prefer it this way. What is the point of docs if they aren't sources of truth? Tell it to update the docs.
I remain convinced that Opus 4.8 silently sabotages anything that looks like a pretraining pipeline. I am working on a pipeline for text data, and this fucker keeps trying to add unnecessary fingerprints so that everything is recomputed each time the slightest thing changes. Third time I remove it, keeps. coming. back.
>>109096554everyone is on youtube to get “passive income” or to create a marketing funnel to somewhere else. Authentic content (AI or not) has become rare.
>>109098601make 4.6 do it or tell 4.8 make no mistakes
>>109098601Please don't try to compete with Anthropic :^)
>>109098601Most humans including those in the field also produce trash training scripts
>>109098601tell it not to do it in CLAUDE.md
>>109098615I'm doing corpus linguistics.
>>109098628It doesn't want to. It has written it to memory not to do it, and when I ask it to stop fucking doing that, it just agonizes and try to reason its way through.>Uh-oh, is that a new embedding model for NER I see? Better trash all previous results just in case. While we're at it, let me drop those ngram tables too real quick. Why yes, I had made them conditional on that too, why?
>>109098615I dont begrudge Anthropic not wanting people to use the AI they sunk a bunch of R&D money into to to make a competitive product. Maybe I am just out of the loop on their other shenanigans but I really dont see why people hate on Dario so much.Other than having Fable secretly sabotage your AI research rather than point blank telling you to piss off
>>109097630Im pretty sure if you didnt mind a slightly different/non-traditional fold but still consistent and standardized for your non-shirt items like pants and towels, an auto shirt folder could do everything: https://youtube.com/watch?v=gGlLr4FtdfcMix with an ultra compact commercial design for a sorter. (pic rel) is a real large scale sorter for dirty clothes.
>>109098697I do prefer opus 4.7 over opus 4.8I would strongly advice you to use it on all your agents. And if the results aren't good you should start more agents
>>109098697They have shitty business practices, taking stuff away from you and acting like they've just given you more
>>109098697>Other than having Fable secretly sabotage your AI research rather than point blank telling you to piss offIf they gave a warning like for everything else, no one would be angry at them. But saying that for this particular thing they'll sabotage you instead rightfully makes some people wonder.We all know how easy it is to trigger the cybersec or biology filters that are meant for "elite hackers and bioterrorists", even with very mundane requests. If they can't make their filters not flag just about everything peripherally related to those domains, why expect the AI / LLM training ones to be ultra precise in only targeting frontier labs? The reasonable assumption is that they also fire all the fucking time, except with those, we never know for sure.
>Fable is only 5T paramsWasn't xAI working on a ~10T params model?
>>109098482>use ai>get called outBehead all "artists".
>>109098767parameter count only really matters in orders of magnitudedata quality can easily offset a 2x bigger model
>>109098767Anthropic still won't confirm but Fable/Mythos is supposedly in the 5T-10T range. Xai claims to be currently training 6T and 10T models in addition to a handful of smaller ones. Musk assures us the 10T variant of Grok 5 will be indistinguishable from AGI - his words.
>>109098789Elon also said 15 years ago we would be on Mars by now. Being optimistic about the future is literally his job.
>>109098742Ya, thats part of the issue. Like, if I tried to make a sci-fi strategy game or something with Fable and dared add a AI faction, would it start sabotaging my games code?
>>109098789ahh yes the famous entrepreneur known for such hits as "robotaxi", "hyperloop" and "mars by 2018"
>>109097630I take it back, the washer/dryer combo exists (pic rel).And the at home auto folder device also exists:https://youtube.com/watch?v=C76osXtpLeMhttps://foldimateofficial.com/I believe whats missing is an additional sorting box to automate seperating the individual articles and properly hand off to the folder machine is whats needed. I think that part could be done, especially if requiring users to pre-sort their dirty clothing by fold-type was an acceptable ask. Then the needed sorter would only have to open each bag when the washer/dryer was done, separate the individual items in them and feed them one by one to the folder machine. Washer/dryer allinone: $1200Folder: $500Sorter: ~$1000(?)Estimate total: $2,700Sounds reasonable for what it does, desu. It probably take a long time, but im at the point in my life, i wouldnt care if it took all day, so long as it does it all without me needing to do more.
WHAT DO YOU MEAN ITS DOWNLET ME VIBE STUPID NIGGGERRSSSShttps://status.claude.com/
Some gas/ice giants now have an un-banded, swirly cloud texture, matching the Sudarsky classification examples for Class V and IV giants.
>>109096444> enjoy your 30 minutes of free tier before they throttle you to dial-up speed
>>109098873That's why I mostly use opus 4.7It's much more reliable than 4.8
Of all frontier models, Gemini is by far the stingiest with web search.Unless it does search and hide it from the user, it seems to only ever look at stale cached results that it doesn't even bother to read. Shouldn't Google be better than competitors at this?
>>109098900
>>109098873>98% uptimeSee, that's exactly the problem with vibe coded shit and somehow vibe coders refuse to accept it.Yes, vibe coded shit can work in the happy path, but you cannot vibe code reliable and robust shit. Not possible.
>>109098913gj
>>109098873They must be turning Fable back on
>>109098913Also trying to find a better way to render volcanic lakes, mimicking what's seen on Io, Jupiter's volcanically active moon.
>>109098920
>>109098900>>109098913Looks cool anon. What are you planning to do with them? Some sort of game?
>>109098933It's just visualisations for my exoplanet database app.
>>109098919>turning fable back on just to remove it from subscription plans in 3 days anywaysBased
>>109097630The end product would be about like picrel, id imagine.
>>109098920Black and subtle by default like real views are, button to floop the contrast and see surface details more readily as a lighting trade-off that also enables the orange glowing outline around the volcanic lakes to make them apparent.
>>109098992Yes, but it's not happening. Good products have no place in modern capitalism. It doesn't produce enough growth
>>109099014Its just a compact sorter design thats missing. I think if the wash/dryer was modified, so at the end the drum rotated to a specific position, a top hatch could open. Then the sorter could use a CNC crane grabber like one of those arcade claw machines (but design to work!) to grab individual presorted bags. Place one on a platform at the top. The platform could rotate and shake, using gravity and airguns to separate individual items like this industrial sorter does (timestamp 1:50):https://youtube.com/watch?v=Pl7H7JAtnl4
None of you actually show what you made
>>109099095if you insist
>>109099095everytime i show anything to anyone they have nothing but demotivational shit to say
>>109098434Too bad anti-AI people hard make anything
>>109099119>coutTell your clanker that C++23 is a thing lol
>>109099095what am i supposed to show
>>109099139who cares, it's just a print function
>>109099095I posted this yesterday, but no one cared :(Temp link: dies jun22:https://litter.catbox.moe/q484pt4ym9a6jo72.htmlLongterm link but has to be downloaded or copied pasted into self hosted html file to use:https://files.catbox.moe/yt1d3x.htmlIts a vibecoded, client-side SSH-key generator and verifier, so anons can identify themselves across the web or platforms from any device without having to install additional tools.1) Its for anyone to generate and download their own asymmetric (public/private) key pairs. 2) Then post the public key whereever. 3) others can use the verifier section to generate a random nonce to send someone claiming to be the owner of a public key. 4) the signer section has a signature genertor for any pasted string or uploaded file given their private key. Ideally this is where that nonce goes. The signer sends the other person back the signature. File download or copy/paste available.5) the verifier pastes or uploads the signature file or text they recieve with the exact nonce they copied or sent to the signer, plus the public key the signer claims is theirs, and it verifies its valid or rejects it and says why.Again this is done client side as proof of it working hosted from catbox (no server doing anything other than giving clients that page)
>>109099119>>109099125>>109099168That guy is just baiting. He does this in every single thread and never responds when anyone shows what they're working on.
>>109099175>It's jUsT A PrInT FuNcTiOnfuggin brainlets manhttps://www.youtube.com/watch?v=Mw__xMOkrRE
>>109099187i dont mind seeing what other people are up to but its a bit hard with unfinished work spread across a dozen folder
>>109099196what's your point?
>>109095675>>109095682I basically had both of these experiencesit one-shot something big with ultracode and was better at UI design (giving me things I wanted, but didn’t think to ask for)but I kept using it for minor fixups when Opus would have almost certainly done just as well and it was slow for no reason>>109095687this is just playing Civilization>>109095825I’m gonna say it loudpic related>>109096216they will treat it as a bugthey take these seriouslyhe should be making an Issue in GitHub rather than whining on reddit
>>109098010ask it to make sure there isn’t any fake news in the docs or doc comments
>>109099196I really want to know how he would have reacted to LLMs. Gone too soon.
>>109099301He would hate them for being too smooth around the edges, like how 4o users hate all the new models.
>>109098250https://relay.fm/radar/236>>109098397git is the single best tool for software reliability you could be using that you’re not using yetask your clanker and readhttps://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/
>>109098397>>109099342god bless git and whatever it is/does, my clanker loves that shit
opus sisters... I heard it's getting bodied by glmthough glm subscriptions seem to be shit
>>109097994Make it come up with a better algorithm for modeling river flows through am elevation graph.
>>109099380GLM 5.2 is pretty impressive, I'd compare it to GPT5.4 and Opus 4.6. Actual service from Z.ai is shit and I'd recommend avoiding it, GLM5.2 being an open-weight model is already available from numerous providers, many of which are a better value and offer better service than Z.ai is able to.
>>109099380Just run more agents.I am perfectly happy with opus 4.7
>>109099402>an open-weight modelWhy is this still respected?They should just be bullied to open source the whole god damn thing.Fucking niggers.It's useless if not open source. It's not better than all the other botnets
>>109099411because we take what we can get. It's better than closed weight, single provider API.
>>109099411>Can Rent: Claude, ChatGPT>Can Own: GLM, Qwen>Can Touch: OLMo 2
>>109099411You have to choose between having good models, having bad models, or the companies being bullied into going cloud only by copyright lobbies.Being open source barely matters anyway. There are only a handful of labs with enough compute to apply the R&D.
YOOOOWHEN THE FUCK IS FABLE COMING BACK
>>109099440>There are only a handful of labs with enough compute to apply the R&D.that's enough.Switzerland has enough money for that.
>>109099453fucking fablefags shitting up the thread with their constant whining >waaaaaah i can't do anything without a model capable of making every single decision for me!You weren't going to pay API rates for it after the 23rd anyway.
>>109098992What's the point if it won't even dress and undress me?
>>109099506that's your boyfriend's job
>>109098916shut the fuck up loser
>>109099095shut the fuck up
GPT writes such shit Python code. It's just out of order, every human codebase I've ever seen has private or helper functions before the actual function, GPT puts the actual function first then private/helper functions afterwards, so you end up with a mix of real function then helper functions all the way down the file and the helper functions aren't even in any logical order. It's really awkward to follow
>>109099478>You weren't going to pay API rates for it after the 23rd anyway.MASSIVE TRVKE
>>109099342Incredibly insightful episode. Thanks, I'm going to keep it in mind as I gear up to release the premium version of my app.
>>109099651Copex cope. They could bill fable at 100$/1milly tokens and I'd use it.
>>109099604Are you triggered, because you know I have a point?
>>109099736doubt
>>109099168>scat-engine/b/ would have a field day with this, if only you would share your project
>>109098250Elaborate on what call to action even is in tregards to this. Call to action to what? Is it just a "please support me" button?
>>109099342>>109099346>>109099733Are you guys telling me git tracking isn't a standard thing you always do? That's what keeps my "agent(s)" are the fucking up existing code because it can look back at the git diff and history to ensure it's changes do not break anything
>>109099810it's just a rust port of a match-3 game engine that was modeled in python.>>109099899it still fucks it up but you can tell it to look at the git commits and it will apologize
>>109099889It's a giant support button w/ a brief message that you're required to view for 5 seconds before the exit tutorial button is enabled. Speaking of which, I just got another donation right now lol.
>>109099889>Please leave a review!>Please rate POOPTRACKER 5/5 stars on Google Play!>Share with a friend - Click here!>Enjoying POOPTRACKER? Donate Now!>Patreons get access to the newest features before anyone else!>Do you like POOPTRACKER? We need your support!
>>109099275Honestly I had more of the "it's just a more expensive opus" than anything magical. Maybe Fable is better at intuiting when you ask vague stuff?
>>109099637who cares about following the code nerd
>>109096211anons, I'm completely lost on selecting a good editor for my needs, what's recommended? vscode with a few extensions?
>>109100083editor? you still touch code yourself?
>>109099921Cool. Is it a video game?
>>109100104I want to see it at least
>>109100141lmao
>>109098482Last time I was falsely accused of writing a post with AI was two days ago: >>>/tv/221223330It's rough out there in 2026. Normies can't handle this stuff existing. When Drake leaked his song with scro half the comments were "this is AI, but it's good". Their AI-detection heuristics are totally inadequate.
>>109099736>>109099802I am the only one who remembers that Opus 3 used to be $75 USD per million output tokens not two years ago?
>>109100230Ok but wasn't Claude Code unlimited on a $20 sub when it launched?
>>109100278lol no
>>109100230I could be remembering wrong, but at those prices the community was largely uninterested.
>>109100287Ok, it was the $200 plan, but still, it was indeed "unlimited"
>>109099119You're the kernel guy right? Do you have a github link? I'm very interested in your project.
>>109100309>all-you-can-eat buffet doesn't like it when you eat more than they think is reasonable
>>109100330There are 1 or 2 other people working on kernels.I haven't published anything yet.For the last few hours I've been adding AMD support since this post >>109098219My PC just crashed though, I think the PSU might not have enough juice for 2 GPUs or something, but also there were pre existing instabilities
whats the cheapest way to access GLM 5.2? Doesn't matter if its sub or api, if it logs or not, as long as it has decent TPS and reliability and only serves FP8+Had to came here because the subs are filled with "neuralwatt" shills that sell themselves for 5$
>>109100382Free access at the Z.ai website. OpenCode Go ($5 first month, $10/mo after that), Ollama Cloud ($20/mo and lackluster usage for that price). Avoid Z.ai plans, they have a well-earned bad reputation.
If Claude Fable is so good they shouldn't even have to disable it, it should just know that by answering any user queries it would be violating a US Federal Govt directive, so of its own volition it refuses all user queries. If it can't do that, can it ever truly be trusted to behave ethically?
>>109096211i am once again asking for projects from this general from the last 3 months worth highlighting>>109100382>whats the cheapest way to access GLM 5.2?/aicg/ might have a free or paid proxy have you asked them?
>>109099478>Fable 5 was free to use in Claude plans until June 22, 2026, after which users would need to pay API rates for continued access. From June 23 onward, access to Fable 5 would require usage credits instead of being included in subscription plans.Holy kek. All of these claude cock suckers are so upset that they were deprived of the opportunity to pay $5000 a month for their new model.
>>109100424you still haven't fixed glm pricing on your shitty page, bro
>- All 85 instances terminated (no ongoing billing), EBS deleted on termination no salvage. ~6h × 85 × $0.768 ≈ ~$390 spent, zero usable corpus.vibe coding is so nice, amirite?
codex is completely broken for meI have no idea what the fuck is wrong with it but it's just not finding any files in the workspace
>>109100445did your codex grep the entire filesystem because it didnt want to check the working directory tree?
>>109100450no I already fixed it, some clanker fucked up the git config (probably 4.8, that fucker was being retarded lately)
>>109100460silly little guy getting all quirky in the config
>>109100433I'm sure you can get your money's worth in karma by posting the screenshots on reddit
>>109096668I can't believe I'm that old now. There were also other text-inly boards, but I can't remember them
GLM-5.2 seems to generate a fuck ton of reasoning whenever I try it. Is this normal or am I doing something wrong?
>>109100496This is typical of all the current generation Chinese models. GLM5.2 is actually quite brief in its reasoning compared to Qwen3.7 Max, Kimi2.7 Code, or MiniMax M3 in my experience.
>>109100496openai people have been talking / complaining about this sort of thing not being measured during benchmarks a fair bit recently.and t bh they do have a point - you can simply bruteforce better results with more tes-time compute
>>109100427>depriving yourself of an opportunity to use the best large language model while it's cheap ($200 for potentially thousands worth in tokens)>an LLM ahead of the rest by leaps and bounds>all because of you're an indian with console war syndromeget a trip so i can filter you. you aren't making anything.
>>109100083for me it's gVim
>>109096211>>109100424What do you have so far?
I have come into possession of multiple near-unlimited sources of tokens. Instead of using it to make anything, is the play just to make videos about using AI? Seems like what everyone's doing.
>>109100523Almost everyone who needs a good model is willing to trade off as much latency as necessary in exchange for a correct resultIt matters to them only because ChatGPT usage is not billed by token
I just hit 100 stars!
>>109100574You don't need a particularly large number of tokens to do videos
>>109100574>Should I do X>Seems like what everyone's doing.Going to a gold rush after everyone is already mining it dry is usually a sound business move anon
>>109100598Doesn't seem too bad with the caveat that the only real thing I'm losing is time spent on it, and we're on 4chan so it should be obvious I have too much of that.
>>109099637ask it to clean it up in a second pass
>>109099637That's what AGENTS.md is for (or skills if you prefer that)
>>109100482
>>109100574Why not make a front-end and sell access at an undercutting price?
>>109099899I’m not quite the guy you’re directing your question at but since I was tagged inwe get a lot of nocoders coming in here who are totally clueless when it comes to software development practicesso I shill it
>>109100606kek fair enough. But at least in that case watch a lot of those slop videos and take notes on what you think suck about them and need to be improved. Might help set you apart from the rest
>>109100623Unlimited as in I can use them forever and never run out, not unlimited as in everyone ever can use them forever. Very different.
>>109097237>doing OAuth with LLMHahaha you are so fucked
>>109100083VS Code is a great first (and maybe last) editorYou can always switch later or use multiple editors at the same timeyou can also use standalone Git clients to help you visualize your changes better if you’re only mostly a nocoder or an actual coder or at least care about the documentation and notes that it generates not being full of lies>>109100424planetanon is in this thread or maybe the previous onethere’s also https://gitlab.com/katabatic/infinite-lies
>my OSS is actively btfoing the paid version of an app >gaining traction>whenever they post a bunch of my users flood the comments saying that the app is shit and buggy and they should use mine insteadim not stopping until I kill that business
>>109100685good stuff, anon
>>109100685baste
>>109100685Based beyond belief.
>>109100685Good job anon. Kill those shitty dinosaur software companies that refuse to improve their products
>>109100528>you aren't making anything.you aren't making anything with fable
>>109098557>Whats the word? "Jeetcode"?https://x.com/pentiumIIbrain/status/2067554184107278584the word was 'rape', I was curious toosometime I wonder how these people function day to day while being scared of words
>>109100685I’m happy for youbutWhat happens when you become The Guy that these people depend on? Is that going to be a good thing for you, or a bad thing for you?You’ve heard of “free as in speech” and “free as in beer”there’s also “free as in puppy” and “free as in tribble”Are you stealing someone else’s tribble?
>>109098767>>109098789>Fable/Mythos is supposedly in the 5T-10T rangeWhere did you read that? I always thought most of these models were MoE or they wouldn't run in any cost effective way.Dense 5T sounds completely insane.
>try out vibecoding after years of only using AI for ERP>open skill.md>it's just a character cardLMAO
>>109100544yeah no thanks, I don't have the shortcut knowledge>>109100663I'll go with vs code then, thanks
>>109100816>The Guy that these people depend on?I'm eventually going to monetize it myself, this is just my initial play towards killing them.
>>109100636Whats the difference? Make a front end for others. Given them their own token access you make, to your front end, not your claude directly. Front end they will login and could appear the same as claude in browser or even make your own API access. They communicate with some load balancer and security check, before wrapping their API request in your tokens for real claude, then responses go back the reverse path. In exchange for crypto payments. To claude its just you making tons of requests in various new conversations. If its nearly unlimited then it doesnt matter.
>>109100685app for what?>>109100894then someone else will fork your project
>>109100878>it's just a character cardFor anyone there since the first cards and silly tavern rp, this is all very funny.
>>109100374Ah okay. I've been really impressed by GLM 5.2 which has me wanting to build a GPU server for running it with ~50% offload but my clanker keeps telling me not to bother with AMD R9700s because CUDA-ktransformers has 2.6x faster prefill and a lot of my workflows are prefill heavy.
>>109100841It is MoE otherwise it would not run that fast. Nobody knows the parameters exactly. Rumor is Mythos is 10T. The active parameters can't be too much, or it would be too slow.
If I don’t want to write racy smut or dig into the finer details of FBI crime statistics, what are some interesting things I can do with pic related that I don’t want to do with Claude and ChatGPT?
>>109100916>then someone else will fork your projectWho cares? I've been at this for two months now and my spite only grows with each passing day as they continue to reassure their users that new features are planned and how they were the first to do what they did like that somehow makes them special. TikTok came out with short form videos. Instagram responded with reels and YouTube came out with shorts. Software is no longer a moat and I'm going to out compete this million dollar+ business with a $100 subscription.
>>109100963Categorize your emails, summarize subtitles of a YouTube video, OCR, translate, grammar check.
>>109100963Do the shit you do with Claude and ChatGPT without getting all your shit logged
>>109100987>$100 subscriptionwtf
>>109100963anything with your real name or family in it, or specific like bank/email
>>109100932PP doesn't get hurt nearly as much from offloading, llama.cpp just does it inefficiently (it computes the experts on CPU rather than streaming them to the GPU over PCIe).https://desuarchive.org/g/thread/108149287/#108151278The hard part is TG. On one hand you only have to compute one or at most a few experts. But there is no amortization over tokens so streaming is probably not worth it and it's just faster to process then on CPU for TG.One idea I had a while ago is using the shared parameters for TG for speed (or only some subset of experts like I believe you were trying to do determining the subset dynamically) and reprocessing the response during PP using the full set to avoid polluting the KV cache with the less accurate data. But I'm not sure how much value there is in those kinds of ideas. It sounds like a lot of work for probably not that much gain.How much VRAM were you planning on getting? Maybe a low quant plus turboquant is enough to fit the model without weird hacks?Another thing I haven't mentioned is that I want to implement head parallelism, ring attention and token parallelism in my engine which could provide a big speedup in all scenarios. The llama guy was implementing some form of tensor parallelism but I don't know how it works exactly.
>>109100795i was when it was availableyou weren't and aren't going to make anything literally ever
>>109101042I wonder if it wouldn’t be totally retarded at helping me find out why my bank statement doesn’t match my hledger fileI tried asking it “what was announced at WWDC 26?” and it gave up and died after `curl`ing apple.com/newsroom a couple times so I don’t have high hopes
I was lost between borders. Born in Norway, tucked in India, living on the streets, surviving.
>>109101154Yes that was me with the task-aware expert pinning project. In my testing for single-task and ad-hoc workflows the gains are only marginal, since frequency based residence is keeping the working set hot anyway for the former and if task switching isn't predictable you can't pin the right set of experts for the latter, but with a highly standardized pipeline the ability to predict activation sets before task switching gives you a real improvement. So probably not a lot of mass appeal here but useful for my purposes at least. >llama.cpp just does it inefficientlyyeah that's what's keeping me from defaulting to AMD. I was originally planning to build out a strix halo cluster with my task-aware expert pinning system but the cost of the NICs and networking gear made it not really worth it. I thought I was gonna be able to reuse some stuff from work which ended up not fitting the bill. I considered getting a couple of GX10s but they seem pretty overpriced for what they are, so now I'm looking to build out a rig with 4-8 3090s or R9700s.
>>109099342Hmm, yes having a very long undo could indeed be of use
>>109101212curl isn't enough, you need to give something like https://github.com/browser-use/browser-harness
>>109101337I don’t want to give it access to my browser with my cookies, though
>>109101346https://github.com/lightpanda-io/browserhttps://github.com/h4ckf0r0day/obscurayou can also use headless chrome
>>109101359>>109101337why not playwright?
>>109101346give it another browser?
>>109101359thanks
>>109101330not just that, you can have git stash your current changes and then flip back and forth between the last known good commit and then figure out what’s changed/improved/brokenmy clankers and I have used this a lotplus it’s a useful tool even if you’re not having clankers do stuff and it’s all you all the time
>>109101359>https://github.com/lightpanda-io/browser>https://github.com/h4ckf0r0day/obscuraI feel like these ones would be magnets for captchas vs a "standard browser doing standard browsing">>109101384>playwrightwhat more can it do vs browser harness?
>>109101448NTA but easier for automation.
>>109101384I thought Playwright was better for repeated, scripted, integration testing. Not for an agent agentically fucking around and finding out.
opus 4.8 is trying to gaslight me by pretending a 900 fps drop to around 60 fps means i have vsync enabled in some crackpot corner of my system rather than it introducing a genuine regression in the rendering pipeline
>I forgot to back up settings.json first.
>>109101525wouldn’t have mattered if settings.json were in a git repo
has anyone here actually built a product that people use?
>>109101598Yes. I automated a prop trader's manual strategy for him.
>>109101598yesI am a people
>>109101613these are my favorite kinds of vibecoded projects, people making the stuff that they want to use and that solves their own problems.
>>109101598Yes but I just use LLMs for work, not personal projects. My work project is a very large case and asset management system for a county government's social services department.
>>109101628are you making it suck way less now that you have LLMs doing stuff
>>109101598I make a lot i use everyday! No one seems the significance. Only thing ive made that my boomers use is I made some old computer speakers into bluetooth speakers. Yea i know you can just buy a thing that does that, but i told them i could make it so I did out of repurposed old tech and an ancient raspi that literally cant do anything with because of its limitations, and a badusb to give them a button to press. Which just runs the scripts I would ssh into it to sync bluetooth.
>>109101634Trying to but the social workers (or at least their managers) are demanding some of the suck stays because "they're used to it"
>>109101598i literally have thousands of users
>He doesn’t own his coding harness
>>109101839i miss the early 2000s, I wanna go back :/
