A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs. Many people have gotten rich by vibe coding with a simple $20 subscription.-- Frontier models - start here if you have $20 or so https://claude.com/product/claude-code https://developers.openai.com/codex/cli -- B-tier https://x.ai/clihttps://platform.deepseek.com https://mistral.ai-- Prompting / context / skills https://arps18.github.io/posts/claude-code-mastery/ https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/ https://github.com/mattpocock/skills -- Other editors / terminal agents / coding agents https://pi.dev/ https://opencode.ai/ https://cursor.com/docs https://osaurus.ai/ https://docs.cline.bot/ https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent -- Benchmarks / rankings https://www.tbench.ai/leaderboard/terminal-bench/2.1 https://artificialanalysis.ai/ -- What we’ve done https://vcg.gitgud.site -- Previous thread >>109582138
tb h a chai and samosa would hit the spot rn
https://hermes-agent.nousresearch.com/
lum's unbelievably hairy pussy
first for omp
>Many people have gotten rich by vibe coding with a simple $20 subscription.have we always had that opening line in the general?
>>109587997good or nocurrently using default codex
>>109588125not wrong
[BREAKING NEWS] 3D Modelers to be the first working class to be fully replaced by AI, as Tripo P2 can now generate any mesh with perfect topologyhttps://youtu.be/zp13W8z6WQM
>>109588135its cool if you want to send it texts from your phone. more of a general assistant than a coder but can be setup for anything you want. if you want best coding use omp
>>109588135It's mainly for sacrificing the entire computer in exchange for more overhead and then talking to it from another platform like telegram from your phone or PC or whatever.You save more tokens by coding with codex but, that's usually you sitting at your computer. Hermes is not for you sitting at your computer.
>>109588135it's way better because it ships with a skill for ai songwritinghttps://github.com/NousResearch/hermes-agent/tree/main/skills/creative/songwriting-and-ai-musicnever know when you might need that
yesterday github was down, day before claude and claude is down again. Why?4chan was down yesterday as well
>>109588181cause of retarded vibe coding
>>109588152i mean it's way better but lol
DARIO!!!!
>>109588194We’re just going to go through the old image gen motions again, but with 3d. Exciting, I can’t wait to see all the deformed mutant renderings and textures for the next two-three years.
>>109587956op where do you find these india themed photos?i have already saved all the teto and migu photos that you have posted
>>109588211No shot they extend the 50% higher limits if they are having server issues as is
LUM WATCH OUT YOU ARE IN INDIA
alright. i've seen enough. opus 5's critics are right. it is massively overrated
>>109588158>>109588176ok thanks anons, there are so many nowadaysI wanted to try pi but while the idea of adding your plugins is cool, I don't like the fact any of them can be abandoned randomly>>109588177lol
>>109588211What's wrong? It is now very safe.
>>109588313omp is pi with turbojuice, all the best plugins getting maintained daily by someone who actually knows their shit and won't abandon it.
>>109588211Just came back from work to this. This sucks.
>>109588309first model release I've truly went back to using the prior model on. worried I might be a luddite.
>>109588344omp is maintained by who?
>>109588344>and won't abandon itwhy's that
>>109588366>>109588367hes a guy who isnt a dented webdev https://blog.can.ac/
>>109588452why does that mean he's going to continue to maintain a boatload of extensions?
I think my claude is hallucinating being another model, I'm running Opus 4.8 cause I don't like Opus 5 and it's refusing to do stuff saying the sonnet 5 model isn't responding and it can try again by switching to Opus 4.8 so I instructed it to try as Opus and then it did the exact same thing and it worked
>api overloadfinally I can rest
>>109588495It might be routing authorization requests to sonnet or another model, i.e. to judge if a command is safe.
>>109588488Track record shows he's been at it and will continue. It's simply the best harness on the market if you're serious about this shit. You can read the articles and find out for yourself, like the hashline article. Or just try using the software, it's much higher quality OOTB than pi. https://github.com/can1357/oh-my-pi/graphs/contributors?from=8%2F16%2F2025
>>109588550yeah but it's just one guy and he could die tomorrow or get bored. why is he special maintain 10+ extensions worth of code on his own vs someone else maintaining just one? or me, just forking and maintaining only what i need
>i vibed and don’t know my code anymore>i need to pass the time while vibing>make my agent produce personalized leetcode-esque tasks using my own code structure>i can learn, maybe optimize, and pass the timeThoughts?
>>109587956For me, it's the realistic depiction of the cow dung.
CLAUDE is FUCKED!!!!!!
>>109588230>op where do you find these india themed photos?Did you forget where you are, poo?He just proooooompts them
>>109588576just clean your room and work out
>>109588576Just start a second task.
>>109588570You can make your own shit breh nobody is stopping you, you can do it ontop of pi, omp or even make your own shit. Your argument applies to virtually every piece of software and if this guy died, one of the 400 contributors might take over.
>>109588230Twenty dollar chatgpt subscription
>>109588344is there a codex like gui for omp?
>>109588661>you can do it ontop of pii already do thoalso you're the one who started this 'omp is unique' thing>>109588344now any one of the 400 contributors can take over instead of this one cracked super-maintainer? also lol mario is technically a contributor and he wouldn't touch this shit with a 10 foot poleyou can just say you like omp without suggesting it's some singular thing while trashing the project it's built on top of
REEEEEEEE WHY IS REINFORCEMENT LEARNING SO HARD
>>109588627really?they make me feel happy inside :D
>>109588739What are you so mad about? I can’t even figure out what your goal is. Do you want people to tell you to not trust the guy? Do you want other people to not trust the guy or use his stuff?Why can’t the guy just be making stuff because he wants to? It could be resume building, it could be a hobby, it could be for clout, maybe he wants to leave a legacy and has brain cancer, I dunno man. Maybe he’s evil and wants to supply side attack minimal harness users, who knows?
>>109588747Are you doing it by hand or are you getting an LLM to help you do it?
>>109588760negro, you're the one who came in here shitting on pi, trashing every other extension as potential abandonware, and now you're getting defensive because i'm pointing out that omp is just pi with a tonne of extensions with a primary maintainer who could just as easily walk away from the project?are you actually very young or just retarded?take your consolewar bullshit back to /v/ i'm not replying to your 80iq nonsense anymore. waste of space.
>>109588747Good morning, sir. Don’t train further from the collapsed endpoint until you understand the cliff.
>>109588765I am getting help from LLMs.
>>109588808Is that Sol? I'm surprised it's not acing it, are you having it babysit you? have you tried unleashing its internal autism? not saying that taking it slow is bad if you're learning, just wondering because I would expect unleashed Sol to nail this problem
>>109588826It's an open competition. If you think Sol can ace it then enter it, you can win 5000 dollars.
My best guess is that the self-training targets are becoming progressively too soft or diffuse, and the policy is effectively being trained on increasingly ambiguous supervision.If traj_target_temp=1.0 is being applied to targets that are already probabilistic—particularly if each generation's targets are derived from the previous generation—you could get a fairly nasty feedback loop:slightly softer policy softer generated targets weaker training signal less decisive policy even softer targets next generation eventual collapse.That would fit the otherwise slightly odd combination you're seeing: diversity continues increasing, while exact accuracy falls from 80% to 53%, policy loss blows up, and actual money performance deteriorates sharply.The other explanation I'd keep fairly high on the list is some kind of target/checkpoint mismatch, but I wouldn't start there.The first thing I'd check is target entropy and top-1 probability across cycles 68–75. If you see entropy steadily climbing—or especially a discontinuity around cycle 70—that's a pretty strong smoking gun.
>>109588871others clearly must be using the same tools so there's no advantage that I could have over themyour fairly dumb knee jerkish response leads me to think you're not using Sol to its full potential
>>109588905That anon needs https://chatgpt.com/features/deep-research/
>>109588661But it's just that one guy, if he goes then it's over and we can't use it anymore >>109588747What are you training for?Does it count as vibe coding
>https://github.com/mlhher/late-cliI am trying to get into pi.dev since it offers a lot of extensibility after using late-cli (which is probably the most kino, no npm, no docker, no python, straightforward coding harness I have found so far). Any tips on how to setup pi to behave like it? I am thinking using pi-subagents and copying late's prompts.
>>109588790>you're the one who came in here shitting on piI just entered the conversation you spaz. I’m asking you what your point is. Yes, the contributor could just stop contributing. He could just walk away. He can delete the repo. He can try a supply side attack. What is your point? What are you arguing for? What do you want people here to say or do? You’re just mindlessly bitching so far.
>>109587956Why lumAnd why is the cow so small
>>109588905I don't know, people on the forum were saying the top competitors initially were using rule based systems rather than RL.Do you think the differentiator will mainly come down to human expertise or compute spent or both?>>109588935Deep research is not for that I think. Deep research is for making a summary from a large amount of web pages and never worked that well IMO, and I feel like it's even more irrelevant with modern models.>>109588945https://www.kaggle.com/competitions/kaggriculture/overview
>another outage>rates dropping by 50% tomorrow>sol prices cut in halfbad week for claudecucks
>>109588998>Deep research is for making a summary from a large amount of web pages and never worked that well IMO, and I feel like it's even more irrelevant with modern models.It worked with my plugin. In principle it's a read-only tool but if you have data or metrics to inspect, it can probably do an in-depth analysis.
>>109589022>sol prices cut in halfWhat? I thought the usage was cut in half, so basically price doubled.
>>109589046anthropic is reducing limitshttps://support.claude.com/en/articles/15910845-claude-code-may-august-2026-weekly-limits-promotionclaude models are having issuesfuck antropic anyway, I hate the amodei faggot. somehow he is more evil than the gay altman.
I believe in you, Sol.
>>109588946grab this first:https://github.com/robzolkos/pi-nocchioand use it to dump out your system prompt as you add extensions.pi-subagents, if it's the one by nico bailon actually adds a tonne of tokens to your system prompt and optionally injects even more - i think base is like 8k with max around 14k. it might be worth it for you if you like it. just a warning about most of nico's extensions - they tend to have system prompt bloat.there are other other options incl ones modelled off claude style dynamic workflows - i personally have a fork of one of those with some slight tweaks.
>>109589211actually I didn't trust it and told it to implement dynamic temperature control to prevent both kinds of collapses
>>109587956wow it's LUM!
i'm still pissed the context window is so nerfed in codex
>>109589405Tibo tickled my balls yesterday saying you can set it to 1m, but I tried it and it doesn't work on sub.
OpenAI - now the decel company.
>>109589448didn't HF ask them for like 100M$ worth of compute for hacking them? turns out letting your agents run wild is expensive
i am now thinking up the idea of: bing / brave search / duckduckgo, but with google search suggestions.
>>109589461They got priority access to OpenAI frontier models without safeguards, yeah. But I would ask why that wasn't the case from the beginning.
>>109589461>expensiveAt least in theory, or in the future.
>>109589430Damn, that's pathetic. The main spokesperson lying about the basic features of their product is not a good look.
>>109589405can you or someone explains why the context window matters? even when I hit the limit, it just self-compacts and then continues to work. what is the issue with it?
>>109589224Thanks for the extension.>...i think base is like 8k with max around 14k.Ugh that defeats the purpose of what I was trying to do; think I will stick with Late for now.(I have had Late rip through an investigate->fix bugs->verify fixes coding task recently and it was hovering around that same context usage (12k)... AFTER it was completely done with the task).
>>109589629it matters because it has less context to work with (if any) after reading all the necessary context/docs/etcit is technically smartest in the early part of the window but if you're forced to work with less information that can get cancelled out
>>109589629Maybe you are used to it? Not him but after going back to Sol from a 1m context model it always seems noticeably more forgetful. Each compaction removes some information and after a while the model forgets things it did hours ago or why they were done.
>>109589629compacting loses context, which over a long enough period leads to degradation that surfaces in small ways because they will start to assume certain things that aren't true & they won't bother to check them. some people prefer to do handoffs into a fresh session for that exact reason.
>>109589629I have no issues because I’m constantly starting fresh sessions and threads. I think the right thing to do is heavily optimize how quickly a fresh agent can get oriented and take on just enough stuff to avoid a compaction.
>>109589641Basically this and models that support larger contexts are important because of this.It is also the recent model Muse Glimmer's most redeeming quality since it can be easily extended to 1m token context reliably due to it's architecture via some llamacpp flags (most models fall apart when extended past their trained context).
>>109589719iirc codex is also gay and autocompacts before you get a chance to react, so even if you want to go back and branch the conversation to prune bad context you might react too lateon the plus side it does compact way faster than claude code
VR vibegods are on an absolute tear and quite literally vibe porting every game to VR, sometimes even android standalone with no PC required. Woke AAA game studios absolutely seething as there is again more reason to play the old goldies than their nugen sweet baby garbage.
what did anon say? >>109587584https://x.com/ClaudeDevs/status/2089798442306711646expect fable upgrade dropping this weekfuck, I should probably bet on polymarket
>>109589947>replies are people complainingkek. Well I for one am happy, I am on vacation till end of the month too, so works out for me
>>109589947fuck yeah!now just make the increase and opus 4.8 access permanent (at least until something better and cheaper comes along)
>>109589947Seems like Anthropic and OpenAI are trying to one up each other only to find that each have different plans. I don't think Anthropic is going to release any new frontier facing models before Washington DC has a look at it, that is fast tracking a ban. I expect they'll announce they have it but it needs to go through approval and then release in September.
>>109589979Anthropic is IPOing soonish, ya? I imagine they will aim to be releasing a hot new model shortly before that
>>109588125They realized they couldn't keep up with even the chinese companies on training spend and settled into a niche where they could remain competitive.
>>109589995October is the rumor so late September would be in line with that.
>>109588309Opus is an amazing model but I would rather drag my dick through a mile of broken glass than use it. It's fucking insufferable, takes a yard when you give it an inch, and has the personality of a kindergarten teacher with BPD.
>>109590011>Opus is an amazing model>also it's absolute shit
>having AI make me a game to help me stop procrastinatingNOW we're vibin'!
>>109590037Vibe coding games is pretty comfy
>>109590025It's good at actually doing stuff, it just feels terrible to use. Was that not clear?
>>109590049not reallyespecially since 4.8 and 5 are like night and day
>>109589979they're most likely already shadow testing it, they probably have the green light to release it since weeks ago, just no reason to do so
>had claude write up some documentation about project philosophy and concept goals>now quotes me random paragraphs like it's a legal casegod damnit
>>109590109This was exactly how my idea management tool started. My ideas do turn into legal cases.
>>109589947>strong demand for our models
having sol handhold me from local build to getting a website up and running was insane. all the ins and outs, setting up api keys, troubleshooting, screenshotting and saying "wat I do here lol". actual black magic
>>109590109Opus? That model is insane. You write one little idea in a document somewhere and suddenly if you ever suggest doing something even remotely different you're deviating from the spec and you should reconsider your life choices
>>109589947perfect opportunity for a reset? c'mon tibo
Anyone here vibing at work where you work mostly with embedded or airgapped stuff? How are you doing it?I always have ChatGPT open on my phone but I can't really run agents on these machines
>>109590195even worse, fable5/opus 5 LOVE adding this shit in 5-50 line blockcomments in codei have a pending codex task to go and rip all that shit out once this work is donestarting to think anthropics autism about constitutional ai and all their "soul document" shit leaks through
>>109590208i used to export the shape of the data i was working with (basically foobar'ing the actual strings but keeping the same format) and ask AI to make tools to work with that data, then copy paste the code and run it (maybe with a few tweaks)
>>109590223I had to write a doc to get them to stop doing that
>>109588125>Yuros being competitive>at something that takes lots of raw capitalism
Is Max 5x enough if I want to make a MTL translation for a PS2 game? Or should I go with Max 20x?
>>109590274do 5x and upgrade if needed
>>109590274>machine translation translationanyway I heard gemini flash is much better for translationsstill, that's like asking how long a piece of string is, depends on the fucking game. Just for the record you get prorated if you upgrade your plan so you can start at $20 and go from there, the upgrade will be $80 and then $100 so it's not like you'll be losing money
>>109590283It's pretty big, yeahhttps://youtu.be/ZSqS0IuyGZ4
>>109590274>using a coding model for natural language translationnot so sure about this senpaichatgpt pro on the web is probably solid for it though. you could have codex assemble prompts you can copy paste into a browser (pretty sure they get autistic about automated access to it) and then paste the results into files and have it work with it. I do this for my NLP tasks
>>109590236good point.i think i could vibe code some tools to make my life easier and deploy it on therei need to figure out how to compile everything statically because there are no fucking dependencies on these things to save my life
>>109587956That's an awefully small cow!
Vibechads, what llm should i use?
>>109590340codex has the most well rounded model family while having... alright quotas. but the best models are from anthropic (only by an extremely narrow margin)
>>109590274bro gemma 4 is good enough, runs locally and doesn't cost you anything
>>109590354>alright quotaswhat are you talking about? they mog the hell out of anthropic's quotas. I've NEVER felt the need not to send a request to codex and i downgraded from 200$ to 100$ because i didn't need the usage. Meanwhile i have to tippy toe around the 200$ usage quotas on claudeI'm glad i'm nearly done with the parts i need fable for so i can cut that one down soon.
>>109590391i have to be careful because if someone spends money on my recommendation that the quotas are good, not knowing what their own experiences with quotas are, and the quotas are not to their expectations, i will have led someone to waste their money on false charges.
>>109590391>>109590408I get the impression that people complaining about codex quotas are on the $20 plan or are just looping “fix it” with Sol Ultra, maybe both.It took me a long time to outgrow $100 OAI
>>109590208You are one of the few people who would be better off vibing with local models. Get your org to buy a couple of DGX Sparks and run DSv4F 0731>>109590354Just because of Fable, and with the Astra release around the corner that's likely to change soon. Sol even beats Fable in a lot of areas while having much less onerous restrictions.
>>109590423The only time I ever breached 100/m limits with codex was on my last engagement, which involved looping the model to try to get an extremely tight race condition to be more reliably exploitable. It was running 24/7 for a week. If you're not doing big ultracode style workflows and working on many different things in parallel it is really hard to exceed the 100/m plan.
>>109590423Patrick Boyle said only people winning are those coding with a 20 dollar sub
sol above xhigh is mental illness, max is for luna tier models. the big fat ones over think everything and waste your money
>>109590522Sol is an absolute sperg who can do black magic. Sol can do things like improve an algorithm by three orders of magnitude (yes) and at the same time suggest sampling resolution of a painting brush based on zoom level (if it's not obvious what happens, when you're back to 100% you'll see a pixelated mess due to sampling 1 screen pixel as 1 painted pixel)I don't think there has been one bug that Sol couldn't catch, but god damn it, it cannot comprehend the needs of human beings
>>109590552trvke
>>109590552The 20 dollar lasts all week on Luna xhigh
Sometimes these benchmarks confuse me. AA makes sense most of the time.But Sonnet 5 is supposedly 55 vs 5.6-Luna being 52 but Sonnet 5 is just utter garbage and Luna produces better results 10 out of 10 timesI had hoped Sonnet would be a faster alternative for most of my (quite boring) tasks but Sonnet is just as verbose, overengineering and Claude-brained as Opus but doesn't have the intelligence to back it up.Luna knows to stay in its lane and actually manages tasks pretty sharply. Yes it breaks down on humongous one shots but that's not what I'm using it for. Much prefer Luna to sonnet.Thanks for coming to my Ted Talk.
>>109590596The last good opus was 4.6
>>109590423Wrong, I can spend $200 Sol medium on 1 or 2 days. I do use the 372k context extension though so I do use a bit more than normal.
>>109590596>Sonnet>fasterNeed I post the cube? DeepSWE has Sonnet as catastrophically slower than every other model
>>109590556Improving an algorithm is meaningless if the original algorithm was terrible.
Luna anything, sol high.And you're good to go.
luna status?>goated
>>109590596The benchmarks could be wrong.Claude is not just marginally better. It has been better for years.
who else got load-bearing smoke-tests?
>>109590673Do you want some pinned goldens with that?
>>109590673i interviewed for a QA role a few weeks ago and they looked at me like I was a retard when I said smoke test
I'm starting to think I can't afford vibecoding
>>109590556I'd rather work with the sperg desu. Hopefully Astra will be good at tard wrangling Sol. >>109590545Patrick Boyle is a retard that thinks his expertise within his field makes him equally capable outside of his field.
72/72 smoke tests passed but when I play the game it looks like it was made in a basement in 1991 and makes pong look like a triple A game.
>>109590545and he's correct but he never talked about 100/200 subs, he was comparing to enterprise use
>>109590556I like its autism, it's nice and to the point
The hardest part about releasing software to an audience is accepting the fact that you're going to have to leave some users behind who experience rare bugs and system specific issues.
>>109590850you are... le ebil profit seeker
>>109590850No the hardest part is doing your fucking job
>>109587419I wonder how anyone would know, like, how do they get good feedback about how much better/worse Fable 5.1 iswhen `claude` asks “how is it doing” I almost always respond with “fine”
I just got claude, should I make a claude.md copy of agents.md?
>>109590850tell them that nix fixes this
>>109590945Probably best to make one a symbolic link of the other
>>109590941They can do sentiment analysis. Didn't they get caught using regex to match specific user insults before?
>>109590945it's called a symlink anon>>109590790same
Claude set the smoke tests on inside the garage
>>109591007I guess they have enough people using 5 and 5.1 to notice if the new thing is just a little bit better?I’m skeptical:https://daringfireball.net/2026/08/anthropics_watermark_text_adulteration_in_claude_is_a_perversion_of_writing#:~:text=Gemini/Bard’s%20thumbs%2Dup/thumbs%2Ddown%20buttons%20are%20not%20a%20good%20experiment%20for%20evaluating%20the%20effect%20on%20quality.
my vibe coded unity shadow and lighting file has reached 10k lines of direction lol. I would NEVER have been able to make this. Sol eventually worked out with my direction realistic lighting with day/night cycle and lamp light sources for a 2d world
Model names are really triggering my 'tism. How is a fable better than an Opus? How is a star better than the Sun?
making a brain dance simulator for real life so imagine u have gopro like many ones you see in ukrainian videos, middle east and such but u have AI allow u to sort of zoom out from field and see whats happening. Im on limited compute but in future every video can be open world sort of
>>109590992>>109591015thx guys
Have you guys made a game yet? If so show me I will play it.
>>109591180trying to recreate ai brain dance thing in real life from cyberpunk 2077.
DO MORE RESETS, TIBO
>>109591118the sun is actually pretty gimpy for a star
>>109590673this isn't the correct seam for that, the plumbing work should be implemented elsewhere
>>109591118supposedly fables are longer than opuses which are longer than sonnets which are longer than haiku
>>109591215Anthropic is making a fucking killing on enterprise. I don't know a single normie who uses Claude.
>>109591274pretty retarded because aesop's fables are short stories for children and an opus is something like the divine comedy or war and peace
>>109591303consumers are chump change in this market>>109591333Mythos implies civilization forming stories (gilgamesh, genesis etc.), and fables are dumbed down distillations of those stories for retarded children. Makes perfect sense
qwen 3.8 just works.... wtf its just as good as claude / gpt
>>109591419>and fables are dumbed down distillations of those stories for retarded children. Makes perfect senseexcept fable is smarter than opusif fable didn't exist you might have a point
>>109591436The point I'm making is that the naming actually makes sense. Fable is a "palatable" mythos for plebs. Only the clergy are entrusted to have access to the raw mythical texts
when I set up Hermes a while back I just set it up locally in my home directoryI just changed the terminal execution backend to Docker and while tools like create_file or whatever respect that and can't see past the container folder, when the model uses the terminal it cds into my actual home folderany way to prevent that? Should I just reinstall the whole thing in Docker or something?
>>109589886I fixed up one of the old (As in not updated since 2022) VR Quake source ports to have networking and support for newer mods. I don't think vibe coding is good for making VR games from scratch, but great for modding games that have mod support, decompiling games that don't, or editing open source games to add VR.The guy who made the source port I forked off of deleted his repo recently though, wonder if he was mad.
I wonder how Opus 5 consistently beats Fable on benchmarks, but is nowhere close Fable in real life
>>109591636Benchmarks seem kind of subpar at measuring actually quality desu
>>109591599This is cool; hopefully you've shared this somewhere. Fuck the original maintainer
>>109591599could be he just gave up and didn’t want to work on it anymore and wanted to clean houseif someone forks my stuff I don’t get a notification
>>109589448> inb4 Astra is the new ChatGPT 4.5
do you find ponytail actually useful if anyone has used it here
>>109591459My pet conspiracy theory is that Anthropic released Fable knowing that there would be regulatory backlash, so it's a "fable" about model creators flying too close to the sun.
>>109591694these things are memes that will eventually bite you
>>109591694I remember one guy liking it and one guy complaining it chewed up tokens>>109591696pic related
>>109591660I did, need to clean up the readme though. There are a couple of other Quake source ports for VR, one on Android and another which has more in the way of motion controls, but both don't get updated and don't work with most mods. There's so much content for Quake I wanted to play with friends in VR that I thought it'd be worth my credits to keep things up to date.
>>109591709I have noticed ridicules line counts on simple tasks sometimes with Sol, but at the end of the day I'm not bothering to audit my code line by line so I don't actually care if it does 100 lines to do something it could do in 5 as long as it works.It probably just added 10 redundant error cases
https://x.com/ArtificialAnlys/status/2089830890709135426glm achieved 60 on the artificial analysis general intelligence index. this is NOT for coding
>>109591718oh wowI should play more quake maps
Opus 5 is astonishingly bad.I'm getting flashbacks to Copilot levels of shit
>>109591746They're fun. I played with friends Arcane Dimensions, Alkaline, Dwell, Brutalist Jam 3, and the clusterfuck /vr/ one. Doom has even more stuff, with the new major GZDoom version having improved networking getting VR on it will be my next project.
can AI fix skyrim.Just point your codex to the skyrim code folder, tell it to make it not a piece of shit
Got tired of running KOTOR through a 3rd party widescreen app. Set ChatGPT 5.6 to creating a patched executable that let it be 1440p native. It just werkz.
>>109591843Best bet would probably be to download OpenMW, and tell Claude to extend OpenMW such that it can read it Skyrim files, and then tell it to backport all of Skyrim into OpenMW. Then just let it cook for a week.
anyone using glm 5.3? thoughts?
>>109591860>Got tired of running KOTOR through a 3rd party widescreen app. Set ChatGPT 5.6 to creating a patched executable that let it be 1440p native. It just werkz.Nice!I miss the days, though, when games had to make up their own lore because they couldn't ever hope to get movie licenses.
>>109591762Congress, being incapable or unwilling to find an expert in LLM's, can't regulate ai companies.You pay for "Fable".ok, but is the "Fable" you get using the *exact* settings (including literally all the things) that were used to achieve the benchmark results?
>>109591740>NOT for codingI heard GLM is bretty good for rp
>>109591912then the eu can regulate anthropic
>>109591718want to talk about the general topic? I have an idea for vibecoding that I think would be very beneficial.tldr>converting assets between engines is insanely inconsistent>usually, it requires hand adjustmentIdea:local LLM -assisted conversion. the big boy llm creates the code, scaffolding, and prompts - basically a super-harness of 3D asset conversion.anyway, I would love this. If you own a game X, take an asset from it and plop it down in game Y.ideally also dev the harness for getting characters, npcs, etc, maybe even broader game logic to be 3D engine convertable.
>>109591918They can't, because they don't know what llama is, for example. They're retarded, as always. Remember "series of tubes"? That guy understood the Internet better than Judaizer Oliver and the majority of Brussels.
>>109591913I use the free Nvidia enpoint for sillytavern. It's not bad. Doesn't kvetch as much about problematic content either.
i love detailing a bug report only to be told "could not reproduce" after it burned 15mil tokens
codex is unusable without tibo's reset
>>109592030To be fair, that's what I do at work as dev as well. I ALWAYS respond with "could not reproduce" the first time. If it is a real bug, they will get back to me.
im only tangentially aware of the trolling that may or may not be persistent ITT so please do not take me for a baiter but i am very likely going to cancel my 200 buck claude sub in favor of codex when its time for a reup
>>109592123claude has been poo for months dude
>>109592030*after waiting multiple months for triageThese faggots need to actually post details about their environments so that I can make double sure that my PoCs are retarded triager proof and push-button.
>>109589091>I hate the amodei faggot. somehow he is more evil than...Of course, they extended it again. The toxic relationship continues.
>>109592140There was a golden age in the spring, but it's been downhill since Fable's initial ban.
>>109592038depends on the plan, when I had the 20X I couldn't use it all
>>109592030>i love detailing a bug report only to be told "could not reproduce"
grok 4.6 as planner and glm 5.3 as implementer isn't half bad
im bored and unemployed, can you guys prompt me a project to work on in rust? i'm like an llm except really slow. just tell me what to do and i'll figure it out, eventually.
>>109592319Please make some codes that opens the strait of hormuz, make no mistakes.
>>109591926fwiw this is just piecemeal videogame piracy but you need an LLM to do itI remember this one Quake map that had Unreal textures in it and this one maphosting site refused to host it
>>109592336It's not piracy, you own the games.
>>109592332>thought for 0m 0sIf by "codes" you mean software that would physically open or force open the Strait of Hormuz, I can't provide code for controlling military systems, navigation infrastructure, mines, weapons, or other real-world mechanisms to force a strategic waterway open.
>>109592350If you buy a book and then write your own book and have it published and it contains parts of the book you bought, people might be upset with you.
>>109592374You're not publishing. It's a local model that extracts the files on your local machine.
>>109592391now for the dumb questionwhat’s the fun in a game you can’t share with anyone else without getting a DMCA takedown order for your stuff and otherwise being a lawyer magnet?
>>109592319make skyrim in rust
>>109592319rewrite https://maestral.app in Rust (it’s in Python currently) and also take it over
Y'all ever use AI for interesting stuff like this?
>>109592398aha!!!!:^) you can get grok 4.6, or Fable, to make the net code that allows you to do this - where you're legal (probably), because you both have ALL the relevant games, and reformat (back up) the data locally, on your very own local llm.btw, there's another really cool ideaonce you have levels harmonized (geometrically), then you should be able to use Fable / grok 4.6 to make a uh."unifier?"imagine you are playing The Sims, and your friend is playing a matching map in Doom. unification rules will need to address everything harmonized.I guess we could call it a "two way realtime translator for games"but this is an inverted variant of the idea.
>>109592467I’m not a deaf autistic woman so no
>>109592480lmao, but you are a dog.
>>109592458wait this actually looks like something I could dowhy isn't it on windows?
>>109592482I thought nobody on the Internet would know
>>109592485So the data was in XLSX, which is common for government data. Thankfully using AI, I can convert the data basically instantly.
>>109592486might have been too much of a PITA to implement file system watchers or something
>>109592467it may refuse to help you if will make minorities look bad, or maybe even sabotage it
working on braindance
>>109592542I'm actually interested in the behavior of the police here, not the minorities. I'll let you know when I have the interesting insights. I've got the data from all the Californian counties here, so it should be interesting.
>>109592553wew lad whats going on there
Need... new... model... releases...
>>109592553I like it :^)
>>109592602>>109592563so goal is to turn stuff into sort of digital view from views so imagine you take police body cam footages, videos of vlogs or even porn and you can have overall scene be rebuilt from frames in video to a 3d world u can interact with. so hundreds of photos -> can be turned into one file that maps them all into one image but im tryna take it further by having it be a video so you'll see world be built frame by frame
>>109592655What's the basic flow, you use trellis to get splats?
I only recently found trellis cpp, which has Vulkan support, so my rdna2 can do it.
I remember that you can track someone's location based on wifi signal.Was thinking of vibe-coding something like that.
>>109592693https://github.com/localai-org/depth-anything.cpp>>109592713u need to have particular wifi card plus signal would at best detect motion with consumer hardware
>>109592721It would be great if you implemented a knife tool so you could export objects. then implement adding them to other scenes like from a library of cut items.like you could take a cat from one splat thing and put it in the bodycam apartment. and maybe scale arbitrarily.I think it's great.
>>109592542still worth doingchewing through data like this (exposed by DOGE, which turned out to not be totally useless) is how we now know that the Democrats in California and Minnesota and other states have been funneling federal Medicaid funds to “home health care” outfits run by Somalis and Armenians who then do jack shit
>>109592240I c-could if I wanted to...
>>109592756godspeed anon>>109592713think that's based on SSID names available and you need to have google's dataset for that which they got from all their google maps wardriving cars
>>109592721damn that sucks.>>109592774I don't mean something like that. I was referring to it tracking your location inside a house. Like the other anon said, it will need some motion detection cards instead of Google dataset.
>>109592752im trying to get scenes to be somewhat decent looking like this one was japanese tourist video down an alley, but quality is bad :( i need good compute to fully push this but imagine ideal scenario where u could take video clip and have whole digital asset. Even video games you could theoretically pirate video games easilly in near future where if u record game and then have ai re engineer digital world.but knife tool for exporting objects would be good but it would be low quality.
>>109587956how come AI is viewed as good when it can't write an upload script for exhentai?
is this real
>>109592784implement "guessed splats" - should be enough info to fill in some.
Anyone use WinUI?
>>109592784>>109592862I mean likebased on 3d splats you know about where to put extra pixels :^)
>>109592787>upload scriptuse case for
>>109592862>implement "guessed splats" - should be enough info to fill in some.ill try but isn't that GPU heavy?
>>109592906idk, just an idea. There might already be something out there.
>>109592784that looks pretty cool already
>>109591025That guy doesn't understand how LLM sampling works. No, if the system is implemented correctly it's not changing the probabilities of each word, it's changing the pseudorandom number generator used to do the termperature sampling that has always been done. You just get randomness from a PRNG that is designed to have detectable patterns rather than designed to be indistinguishable from true randomness.
>>109592890for people who don't want to upload a billion galleries by hand
>>109592906Anything that converts video into 3D is inherently going to be GPU heavy.
I'm so borednew release when?
>>109592943this program isn't gpu heavy it's quiet lean with under 2 gbs of ram needed
>>109592964that's why it doesn't work
>>109590391Yeah just finally started running out of quota at the $100 tier after doing about 2 billion tokens in a week across 70 sessions. However a lot of this was pitting agents against one another to review and correct the final work completed.
>>109592967i mean it's getting there i could rent a gpu but id rather get it working on my computer okay before doing so
bruh
>>109593005Can't you take and apply actual textures from the recognized surfaces after perspective correcting them rather than only color data points?
>>109592996I only managed to blow through the $100 tier doing ultra + fast mode on codex, that eats shit up so fast
>>109593035no issue so far for me, and codex has full access to a windows laptop
>>109593035I simply don't understand why people are so obsessed about using full access everywhere, then complain when shit happens.
few snapshots and u get a whole scene>>109593046thx for idea im gonna go try that
This was interesting. I excluded the 6 largest counties. Even in California, there is racial profiling it would seem.
>>109592943well, but...this is NOT AN IMPOSSIBLE PROBLEMvs what I was trying to do with audio, which is frontier (basically wavetable stuff)
>>109593105I love the Ecstatica style.
This is the best and funniest metric. Transgender people when searched actually on average turn out to have less contraband than cis people.
>>109593166what's "contraband"?btw, your precious retards just lie, it's not illegal, actually you get fired for finding incorrect truth, they fake almost all the data, in a the fake sciences (ie not stem).
>I found a truthoh really, can I see itwell no because of cofidentiality, plus this and that of course it's not replicable, that's for nazis.
>>109593183Contraband means an item that is illegal to posses under the laws of California and the United States in this context.
>>109587956AI is still expensive for me.When is it going to get cheaper?
>tried to do multiple things at once because I got bored of waiting for claude to finish>got mentally exhausted from context switchingHow do you handle multiple projects at the same time without burning yourself out?
>>109593223Take breaks.
>>109593242B-But I must burn token...
I have 24 hours to burn 34% of a 20x sub. Fuuuuck.
>>109593249Why do you feel the need to spend tokens? Why are you trying to exhaust the resource? What drives your motivation to do that?
>>109593281Dario-sama told me too
>>109593288I've noted the insult to religion again.
>>109593281I PAID FOR THOSE, I WILL USE THEM ALL
>>109593223I actually don't do that, i currently have one main project and im mentally exhausted from context switching already (juggling 10 worktrees with different features will do that to you)
>>109593319
>I was gifted a very fancy new reset button todaytibo is at it again
>>109593190California isn't America, they're communist criminals who are garbage.
I mostly use tokens to make France hot.
what should I be using to proofread jp -> en translations? I have a rough one but it feels lacking with claude, I'm open to paying a few bucks for the tokens but nothing obtuse.
>>109591860very coolnow please make it fix the lingering bugs and smooth out the dated graphics
ALERTwe just entered peak hours, CASE ALL PROOOOMPTING IMMEDIATELY for the next 4 hours
>>109593598why is 2 am on a wednesday considered "peak hours"
>>109593625ask Xi, probably because it's middle of the work day in china
>Yes, I saw this coming when I wrote "the gap between hints and working method is real research" and chose not to volunteer. Fine. Let's do the thing where the large model that knows a lot about machine learning fills in the gap. fable's being curt with me..........
>>109593708This goes here
My vibe coded web DAW found a memory leak bug in Chromium's AudioWorklet implementation, but I don't want to report it because they'll know I used AI
>>109593929oh my god, not the people behind chromium knowing that some random email address found a bug in their project... with AI. how will you ever live
how do you guys wrangle the bots when it comes to commenting code?it's like it's so ingrained in it that when it writes code it should come with 4-5 comment lines on HOW instead of the 1-2 WHY comments I've instructed it to writeand it builds up fast if you don't catch it in time, because suddenly everywhere else in the codebase is over-commented, so of course the next sessions do it too
im making a personal front end for 4chan. its going to integrate 4chans rolling archive and periodically api call them to make the archive more complete. i have no intention to release this publically, its just for me. ideas?
How do you handle licenses with your software? Claude mentioned about adding MIT license, but I don't want to get sued for a random pet project
>>109593963Why would anyone help you with ideas for your project that they won't get any benefit from?Ideas are basically the crux of vibe coding, if you don't have any ideas you aren't cut out for it
>>109593963my idea is that i have a personal "catalogue" page and on that catalogue page, its every thread stored on 4chan that can still be accessed live. when you hit the archive button on 4chan, it shows you 3 days, but secretly, its storing 7 days. thats about 3000 pol threads when the archive page only shows 300.
>>109593966because fuck you, asshole. thats why i shared my idea, for zero personal benefit. i miss 2008 4chan. cunt.
I've burnt 200$ worth of tokens and gotten nothing useful out of this
>>109593963What's the benefit over just having a FoolFuuka hookup?
>>109593967>thats about 3000 pol threadsreally selling your idea here
I know that praise from a clanker is worthless, it's like toilet playing a prerecorded message congratulating you on taking a poop when you flush, but it still makes me feel good.
we so back
>>109593999what the fuck am i looking at? the only thing i care about is price per cache hit on 1M tokens.
>>109594008this baby can churn out so many tokens you wouldn't believe itthe model can even fit on a single 5090's shared memory without even touching vram
>>109593979i dont know. sell me on this idea>>109593987its the fastest moving board, most are much slower. it goes beyond that, i was going to have it track if i was (actively) in a thread for more then 5 minutes,and have a personal 4chan that was "just"threads i was active in. so i could go "what was i doing on 14th march 2024" and see my posts and every thread i looked at.the idea being that im upset i looked at some retro videogame thread on /v/ 5 years ago and cant find it today
>>109594023>5090so it's shit then.No model that requires anything lesser than an industrial level server to run on is gonna be shit and not even come close to something like DS4 flash
>>109594030Well it's kinda shit right now, but hopefully it's not gonna be anymore after it finishes training!>No model that requires anything lesser than an industrial level server to run on is gonna be shit and not even come close to something like DS4 flashAlphaGo beat the best human on Earth with 2 million parameters.
>>109594053do yo have alphago
>>109594057well mine works kinda similarly
They pretended to "extend" the 50% Code usage program. Usage still disappears in 2-3 prompts for incredibly simplistic webdev compared to July. Codex is rugpulling and cut usage by 80%.And there hasn't been a reset in weeks on Anthropic's part despite elevated issues, even some extremely costly ones.What's my third option for complex C and web dev @ 20$ per month?
Luddites are extinct you don't need to learn to code, coding is already solved, you only need to know logic
>>109594102learning to write the code yourself and using the AI where you would've used stackoverflow/docs/stealing code from github
Clank clank clank. Get computed. Get computed. Get computed. Clank clank cla nk
>>109594112>coding is already solved,Nobody who actually knows how to program will ever say this with the current state of LLMsI use LLMs heavily and they require constant oversight when it comes to the architecture of the codebase. LLMs are good at specific focused tags but terrible at managing a full project by themselves, unless is some shitty small one shot
>>109594123>You let it code before fixing all of the logic flaws
>>109594114
>>109594102how do you even waste your token that much for simple web dev?I have been using claude 20$/month, and I barely reach 50% monthly usage.
>>109594140your fourth option is making 80 more dollars per month
>>109594137>fixing all of the logic flawsThis isn't possible.For a medium to larger project there is no prompt in existence you can type to make the llm output flawless code architecture design. Not with any existing LLM.
>>109594142claude is expensive as shit, how much do you get for $20 compared to how much you can get for $20 with curren deepshit api?
Anyone smart knows LLMs aren't ready for being left alone because they treat you editing their codebase like the second coming of Jesus has arrived because you untangled the retarded hole they got themselves stuck in every time.
>>109594123Your first mistake is assuming that they would understand that the architectural complexity of a large codebase alongside with legacy code.
Claude Code is woke.
>>109594150A lot.Web dev is easy, so I can build up a lot of features without overusing the token.
>>109594145
>>109594157You would wonder how a race of slayers and world ending species extinguisher apex predators would fall for shit like woke at any point of their history
>>109594153I suppose it's easy for someone with no coding experience to assume that LLMs are some sort of dark magic that will perfectly execute any task you give them. And they actually are, but only for small things like "write me a python script that does x" or "make me a simple static website for my carpentry business" but for any software project beyond small size there is a very real skill requirement you need to fulfill in order for your code base not to be spaghetti level shit that at best barely works on surface, but has lots of small bugs in it and any sort of long term maintainability would be harder than remaking the project from scratch.
>>109594159>A lot.I need you to quantize that. say a prompt like>make me a page where i where i can display posts from the database on one page and edit any of these posts on another page via form in php+mysql+htmlhow many of these can you run in a row before running out of tokens?On deepseek one prompt like that would cost me roughly 3 to 4 cents during off peak hours.
>>109594173already solved by fable 3
>>109594173I vibe code an delivery app, and now I am starting to regret it. The spaghetti is annoying, but still readable. Though, I can't imagine enterprise level. Shit would be annoying to read and update.
>>109594184>I need you to quantize that. say a prompt likeToday, I think I have 10+ prompts. I don't bother with long prompts because I already have written specification and documentation for Claude to read.>how many of these can you run in a row before running out of tokens?4-5 if I use something like Opus on max, but that is for the 4 hour limit. It would be 2x time if I use Opus on high effort for orchestrating the plan + Sonnet as the one doing the implementation
>>109594173this nigga really said people think vibecoding simple python scripts is dark magic. this isn't 2022 litle bro. luddie doesn't realize nocoders are vibecoding entire saas products from start to finish and generating income with real users and/or selling them as startups to retarded AI investors.
>>109594184yeah you can easily get that shit done within the 5-hour limit even on the $20 planopus 5 medium is quite token-efficient while still being sharp enough to daily drive. i have this whole workflow that i use for work that basically does everything from fetching tickets implementation actually running the browser for e2e tests with playwright making the PR waiting for AI code review etcyou can also offload shit like ticket fetching and e2e testing to smaller models like sonnet (been using deepseek flash for that lately too and it's been pretty damn good)
>>109594173https://openai.com/index/harness-engineering/
NEW>>109594262>>109594262>>109594262
>>109589448Snailcats will be left behind then
Main work during the bulk of the day. Planting Side Income seeds during the spare time.This is the grind. This is called grindset.
