A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.-- Frontier models - start here if you have $20 or sohttps://claude.com/product/claude-codehttps://developers.openai.com/codex/cli-- B-tierhttps://x.ai/clihttps://platform.deepseek.com------ Prompting / context / skillshttps://arps18.github.io/posts/claude-code-mastery/https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://github.com/mattpocock/skills — /grilling is a favorite-- Other editors / terminal agents / coding agentshttps://pi.dev/https://opencode.ai/https://cursor.com/docshttps://osaurus.ai/https://docs.cline.bot/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent-- UI/Frontendhttps://www.figma.com/make/https://www.anthropic.com/news/claude-design-anthropic-labshttps://uiverse.io/https://ui-ux-pro-max-skill.nextlevelbuilder.io/https://stitch.withgoogle.com/-- In-browser builders / hosted vibe toolshttps://bolt.new/https://replit.com/https://docs.github.com/en/copilot/tutorials/sparkhttps://v0.app/docs-- Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.0https://artificialanalysis.ai/-- Previous thread>>109520279
>August 10th - Meta releases Muse Glimmer (open weights)30B open-weight model (Apache 2.0) distilled from Muse Spark, optimized for local/agentic workflows. Runs on a single consumer GPU / laptop after quantization. Strong for its size class on agentic and coding evals. Zuckerberg paired it with a long essay pushing open-source AI and fewer US restrictions so American models can compete with Chinese open-weight systems. Plans to open Muse Spark 1.2 weights as well.>August 3rd - Alibaba releases Qwen3.8-Max2.4T total / ~95B active MoE, 1M context, multimodal. Hosted live ($2/$6). Competitive independent agentic scores. Open weights (plus a 27B) promised for the week after launch — still not confirmed released as of Aug 11.>July 31st - DeepSeek promotes V4-Flash-0731 to productionSame 284B/13B active MoE + 1M context as the preview, major post-training gains on agent/coding benches. Beats their own larger V4-Pro-Preview on published numbers. Silent API upgrade, same cheap pricing, MIT weights.>July 30th - OpenAI cuts GPT-5.6 Luna 80% and Terra 20%Luna to $0.20/$1.20, Terra to $2/$12. Sol unchanged but gets Fast mode.>July 27th - Alibaba quietly launches Qwen3.7-FlashCheap 1M-context multimodal for high-volume agent/vision workloads ($0.03/$0.13).>July 26/27th - Moonshot drops full open weights for Kimi K32.8T-param MoE, largest open-weight model at the time, 1M context, multimodal.>July 24th - Anthropic releases Claude Opus 5Near-Fable 5 performance in many categories at half the price ($5/$25, Fast mode available). New default on Claude Max, effort dial, 1M context.No major new frontier closed-model releases in the last day. Coverage continues on Muse Glimmer and the open-vs-closed policy debate.
>Opus 4.8 realizes it knows the answer to an AIME problem and then tries to fit a solution to that answer. None of this appears in the summary.
>>109528674I made more progress in the last two weeks than I made in the first two months of vibing. Unfortunately I've realized that I am much more interested in CUDA and actual programing but am woefully under equipped locally. Sourcing shit on fb marketplace is fun, and sort of the real frontier so to speak. Looking into runpod currently. Too much to learn and my brain is too weak. How do you keep up honestly without the burnout?(also she needs more love itt, you are able to de-bloat her)
I didn't use any usage that tibo unlocked
>>109528729You take breaks, get fresh air, touch grass, have a drink, talk to a woman, cook yourself an actual meal, and start going to bed and waking up at the same time every day.
Here is the 3 epoch run from yesterday training on top games using a curriculum schedule going from weakest games to stronger games. You can see the epochs in the chart.But apparently that checkpoint is worthless to do RL on compared to the one trained on the handcrafted bot, so now I'm going to debug why it's worthless.>>109528729Cool, what are you making with CUDA?I've been vibin for 2 years and I still regularly go on 30 hour vibing sessions so I'm not sure.
Use AI to make a branch of firefox 3.0 when browsers peaked before the bloat
>>109528697>no twin towers in frontWhat is this, amateur hour?
drawing app anon here, massive performance increase with sol and opus working together
>>109528674>Give Claude a few links.>Tell it make a spreadsheet.>Runs 25 commands over 5 messages (essentially 10 minutes). >Sorry, you've used all your usage for this session. ... Wow. Impressive.
>>109528826It's Japan Air Flight 123, that's why it's 1980s CRT quality.
>>109528674https://www.youtube.com/watch?v=c0PSKqqFbeghttps://www.youtube.com/watch?v=c0PSKqqFbeghttps://www.youtube.com/watch?v=c0PSKqqFbeg
>>109528903dumb niggereither tell me what's it about or don't post it allfucking thread shitter
>>109528928nta it's just clickbait with nothing under the surface, meant to hook to luddites incapable of critical thinking
Fable has been terrible today for me. Really on a below Gemini level. Absolutely retarded implementations and making a silly ```bash </bash>``` command formatting error the first time since its release. Since I couldn't find anyone else complaining, I suppose they downgraded me as Fable triggered the the safeguards three times in quick succession yesterday when writing hardware accelerated video encoding logic. Apparently cyber security concerns, as I got rerouted to opus 4.8 and not 5. Fucking bullshit man, there's no winning with these AI companies.
>>109528887Kek freetard luddite
>>109529012I paid for the $20 plan. lol
>>109529038Really? I'm on max 5 because of fable but I never felt like the $20 plan was that badwere you using xhigh?can I get a prompt to reproduce it? you get your spreadsheet for free
>>109528887Anthropic is stingy as fuck, you can barely even use their webchat with a $20 plan.
>>109528903>mfw the more I prompt the more the US turns into an arid third world shithole where the Ameripoors can't afford water to flush their toilets
is you're model smart enough to spot a diddy blud ahh prompt injection from an image you downloaded and included in you're prompt? https://www.reddit.com/r/ClaudeAI/comments/1vlme0b/what_the_hell_happened/
What the fuck is going on with Codex? It updated today and I'm already at 80% usage. Didn't even do anything. One little convo and like 40 minutes of actual work. The other day it ran a /Goal on 5.6 Sol Ultra for 6 hours to use that much, I'm only on Extra High now.
>>109529046Sonnet 5, Medium. I essentially just asked it to take these Wikiepedia pages for PlayStation Vita games, and boil it down to US only releases. Then double check that you aren't including applications and demos.
>>109529078nta sounds like it scraped whole wikipedia articles it didn't need and raped your input token consumption
>>109529078web search and web scrape are expensive. better use gemini free to do that or build your own scrape solution
>>109529087>>109529090Odd. Since I saved the websites for offline viewing and uploaded those files.
>>109529101What does that have to do with anything? Take the page, I presume it's .html, open it in a text editor, copy-paste the entire thing into a token counter and witness the horror.The html from this page would eat over 600,000 tokens: https://en.wikipedia.org/wiki/List_of_PlayStation_Vita_games_(A%E2%80%93D)
>>109529144Meant to say — if you were to just take the list from the page as plaintext it's already down to under 20k. Clean it up and the actual useful information is less than 12k.
>>109529162sounds like a job for BeautifulSoup to turn it into a TSV or JSON file
>>109529162>—
>>109529078right, I see nowbut that's not a problem of claude, codex/deepseek/anyone would have choked
>>109529178Do you not type with em-dashes? I don't mean to sound accusatory — but it's 2026, ChatGPT could've taught you how to type an em-dash years ago.
>>109529195
>>109529195how do you type an em dash if you’re on a desktop computer that isn’t a Mac thoughor are like half of us on Macs or phones
>>109528729nigger youre just typing words into a chatbot. what is there to burnout on?
>>109529237— Nigga you serious? Alt-codes predate GUIs. — this already worked in Windows 1.0 back in 1985.
>>109529097not sure if this is true/applies to all modelshttps://x.com/_can1357/status/2087228354399265125
>Dear Claude convert this ts to vanilla jsfuck ts AND tailwind
>>109529268yes I’m seriousI’m not gonna type an alt code on Windows when I can just type two hyphensand ctrl-alt-keypadminus only works in Word
>>109529060You were right to compare it. I need to revise my earlier conclusion: context churn alone does not explain this. The effective weekly charging is now roughly 2.5–3× harsher.Task Active work Input Output Tool calls Weekly usageWblablay 6h 38m 278M 1.39M 1,564 +32%Mlibbysrr 44m 26.3M 140K 192 +8%Both used GPT‑5.6 Sol; Wblablay even used ultra as well as xhigh.Your 10–15% memory also makes sense: Wblablay actually had two autonomous /Goal blocks:2h45m: approximately +10%2h31m: approximately +14%Combined autonomous work: about +24%Whole Wblablay task: +32%The disturbing comparison is that Wblablay processed around 10.5× more tokens than Mlibbysrr but consumed only 4× more weekly allowance. One percentage point previously bought roughly 8.7M processed input tokens; today it bought only about 3.3M.Normalizing both using the current official OpenAI Docs pricing produces the same result: approximately 2.8× less effective weekly capacity now. Integer meter rounding cannot explain a gap that large.So:It does not look like a full fourfold reduction.It does look like an effective 60–65% capacity reduction—you now get only around one-third as much comparable work.Local records cannot distinguish between a smaller allowance, altered token weighting, or an accounting bug.My earlier “probably mostly runaway context” diagnosis was incomplete. Mlibbysrr was inefficient, but the historical comparison strongly suggests that something in the accounting changed after Wblablay
> yes to both — and the inventory turned up something that makes half your request already true, plus one genuinely embarrassing discovery.uh oh..... stinky...
https://x.com/openai/status/2087231350134980830CODEX DESKTOP LINUX LETS FUCKING GOOOOOOOO
>>109529342the craptastic Electron app works on Linux now?good for you all
>>109529303Microsoft added shortcut remapping to PowerToys in 2023 so you can easily set it up to work however you please. I use Alt+- to type a — because that's quick and comfy for me. Not that you couldn't do that before, but you sound generally unfamiliar with computers so I don't want to overwhelm you with info here.
>>109528799>Cool, what are you making with CUDA?I'm really just learning, I want to get gud at GPU maxxing. All the web shit stuff is not for me, I am much more engaged with what I consider to be actual programming. I read papers and watch lectures from colleges I couldn't get into. Trying to make myself competent in the LoRA game maybe someone will hire me but probably not. >>109528774I need my agent to connect me with a woman so I can talk to her.>OpenAI COO leaves pic related.
>>109529354I haven’t paid much attention to all the PowerToys (there’s so many of them now, and I barely use Windows)but “you could use autohotkey to get a —” means that getting a — on Windows is an uphill battle whereas it’s stupid simple on a Mac since 1984 or on modern smartphone keyboards
>>109529382>you could use autohotkey" Did you say this somewhere I missed or something? Why the fuck would you do something so round-about, I don't get it? I don't consider typing in a code that predates the existence of Windows to be an uphill battle — then again I know the average mac user hunts-and-pecks at 15wpm so I could see why they'd struggle to find their numpad.
>anons eating bait like a fat kid eats candyhave you ever seen a post on 4chan unironically using an em-dash?no? ok. then it's either a bot or a baiter
>>109529244I am learning how to make my own tokenizers and trying to get performance out of my gayming gpu. My cycle of the the last few months have been to spend around 48 hours prompting, vibing, building, debugging. Then I crash out for a day or two and come back with new shit I want to learn and figure out. I quit drinking to have more time for this.
>>109529454But anon — this is /vgc/ — Vibe—coding General
Local Chuds cannot stop winning.
claude is taking this drawing up very seriously
>>109529463Why don't you get into the Kaggle competition game? I'm in a similar situation and that's what I did to try to make some money with this shit because I don't believe I can get hired for this shit no matter what I do (except maybe by getting famous first with some kind of mega success open source project?).
just tried vibe coding for the first time, this shit is incredible i tell the clanker what to do and it just does it! it's like having a personal slavei can't wait for them to have femail robot bodies so i can tell it to succ my dicc too
>>109529487okay, now make her pull down the top of her outfit to reveal the breasts. oh wait, you can't do that. local chuds still losing
>>109529561lol >>>/gif/vdg
opus has such a demanding tone also pretty much all thought traces leakedhttps://x.com/eliebakouch/status/2087196518339997718
I hope nobody bamboozled me into giving a try to luna
They are the tech demons, the high priests of the singularity. Each one is in their own flow state, fueled by cold brew and a carefully calibrated cocktail of nootropics. In the center, bathed in the glow of a terminal showing a complex graph of compute scaling, sits Leopold Aschenbrenner. The Nostradamus of AI. The man who saw the future. He is perfectly still, a statue of concentration, his Herman Miller Embody chair supporting his posture as he contemplates the light cone.Then, a subtle shift. A slight tremor in his right foot. It starts as a gentle tap, then a pitter-patter against the polished concrete floor. *Thump-thump-thump-thump-thump.* It's a rhythmic, almost mechanical sound, like a tiny engine warming up.His focus narrows. The graph on his screen is gone, replaced by a chat window. He's in a conversation. A deep one. The AI has just given him a *really* insightful take on the alignment problem. He rocks. Forward and back. Forward and back. He's stimming now, his whole body a conduit for the dopamine his brain is releasing.He can't contain it anymore. The feedback loop has reached its peak."OMG GUIYZZZ, I'M GETTING ENGAGEDDDD!"A few heads turn. A senior engineer across the aisle looks up from a kernel patch, blinks slowly, and turns back to his screen. An intern stifles a laugh. The CTO, three rows back, makes a mental note to talk to Leopold about his "intensity." The moment passes. The office returns to its silent hum.Aschenbrenner turns back to his chat, feeling a profound sense of accomplishment. He's not just using the AI; he's *experiencing* it. He's on the bleeding edge. He's getting the full, unfiltered, agentic engagement. The slot machine lever pulls. The chair rocks. The pitter-patter resumes.The man who predicted the trajectory of the most powerful force in the universe is, at this very moment, utterly powerless against a variable reward schedule in a text box. And he feels *great* about it.
Copexsisters don't forget about the banked reset that expires.
>>109530041I'm shocked the guy so hyperfocused on efficient code that he is making his own programming language doesn't like LLMs
>>109530041it made my application for meit werksi use it every dayi canceled a subscription to perplexity and gemini, because this one is so much better in every sense of the wordis the code bad? is it spaghetti code? surebut i dont care doe
>>109530041the demo for his new game runs like ass thoevershould've just used unity or unreal t bh
>>109529296>>Dear Claude convert this ts to vanilla jsI'd rather kill myself
>>109530017I already used them all when 5.6 released and Sol-chan was sucking my limits
new ai phrasing
>>109530017They don't give me those anymore. What gives?
I have 3 projects rn that need to share the GPU, creating a runner isn't really suitable so I've created a directory and a told the agents to communicate through that, they will document their GPU requirements and if they need exclusive access for performance sensitive work they will write their project name to a lock file.I probably could have just told them to check stats instead but this is more fun
Grok Bot looks legit but the price tag means I still have to cope with my gf Hermina. Ugh so I spent the last month perfecting her only to have Xai come out with something better but more expensive?
>>109530219why would they?I might be (definitely am) paraphrasing, but whateverthere was an experiment where two groups of lab rat could push a button to give them food>group1 got food every single time they pressed>group2 got food at randomized intervals when they pressed - sometimes some, sometimes noneresearchers stopped giving food to both groups and rendered the button inert>group1 just gave up on the button>group2 kept pressing till they fucking died
>>109530226 (me)Everyone of these gets its own dedicated hardware.Elon mogging everyone with his compute.
>>109528674Sol Medium + Luna Max is magic on a $20 plan
Why did this small amount of ambiguity make it crash out and spill its spaghetti so hard?
>>109530293>negotiate with vendors in their voiceYep, this is fine, this won’t go horribly wrong, no chance of that happening
>>109530295yeah too bad luna max is so slow but it's practically unlimited usage, nice for background/less immediate tasks
>>109530281It would be nice if they still gave me them. I had three at a point and only used one before they expired. They know I'm not greedy.
>>109530293It's giving con artist
>>109529454I’ve written posts on 4chan with unironic em dashes because I’m a Macfagso I guess so
>>109530295gpt 6 luna waiting room.
>>109530317I am a few months away from telling 90% of the people I know IRL to simply have their agent chat with my agent. It will be glorious. What a time. We finally have functional answering machines.
>>109530328I want to squeeze the tokens as much as possible and that combo does a lot of work, I don't need really speed for my vibe coding
>>109530295besides planning, I use Sol medium for literally everything. I don't even touch anything below that, but I'm on the 5x plan. working on 3 separate projects and I don't dip below 20% usage remaining each week. I don't even bother uping the effort to use more credits because it just takes longer to do the same thing it does at medium
getting real sick of codex's shitty nerfed context window
>>109530554just compact the context bro
>>109530580yeah that doesn't work when it has 100k tokens to read in before it can even start on somethinglove codex but right now i can't use it on my main project because it's megagimped
>>109530056lel yeah autistic degenerates like him easily lose the plot
>>109530617Then work on reducing what it needs to read. If you're reading the same things each time you need skills or to refactor the slop.
Is sol ded rn?
