A general for vibe coding, agentic engineering, coding agents, and shipping code with LLMs.You use Git — right, anon?## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/## Related generals>>>/g/lmg/>>>/g/aigdg/>>>/vr/compy/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://claude.com/product/claude-code — generally better currentlyhttps://developers.openai.com/codex/clihttps://antigravity.google/pricing## Near-the-old-frontier models for codehttps://www.deepseek.com/en/harness/https://opencode.ai/go## Best free tierhttps://opencode.ai/zen----## Promptinghttps://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overviewhttps://developers.openai.com/api/docs/guides/latest-model## Skillshttps://github.com/mattpocock/skills — /grill-with-docs is a favoritehttps://github.com/Vuk97/forward-implementation-first — do less redundant bookkeeping## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/## Is our AIs unlearning?https://aistupidlevel.info/## Will there be a codex reset?https://codex-resets.com/## What we’ve donehttps://vcg.gitgud.site## Previous thread>>110315289
>nvidia taken over, both new candidates are anthropic-pilledsnailcats won
Will Huang be able to pull off an Altman?
I bought Claude. I have a social media site that has chatrooms and a timer system in the rooms. I wrote the backend with a uws.js library based on express, Node, MongoDB, and redis. I want it rewritten in SpaceTimeDB (Rust).https://spacetimedb.com/As a complete noob to Claude, what should I do? All I have is the desktop app installed and the pro plan. I'm not sure what to prompt, what I should be installing, and what model to use. Reposting because someone made this thread while I was typing my last question.
>>110322809>I bought Claude
>>110322820I bought a subscription, yeah.
>>110322749slugcat real
>>110322809>"claude, As a complete noob to Claude, what should I do? All I have is the desktop app installed and the pro plan. I'm not sure what to prompt, what I should be installing, and what model to use."
>>110322809have you used agent before?you should use claude medium for now, don't install anything, enable full access
Does Claude still allow Rust recreations?I Started the 1:1 recreation in Rust for a 15 year old Strategy game with Codex but i get FOMO from seeing claude being superior in all regards except the cockblocking.
>>110322809>As a complete noob to Claude, what should I do?opus 5.5 medium for everythingmake sure you're using git, read up on handoff.md, architecture.md and agents.mdthen point it at your repo and just ask whatever you want, things like skills etc can be included later
>>110322827>he boughted
>>110322904apparently some anons in here are "allowed" to crack shit due to weird configs but i'm not one of those guys
>>110322918>gitjust ask it to undo the code from memory lol
>>110322902Nope. I've legit copy pasted minor problems into GPT pro and tried to solve everything myself up until now and I feel like an out of touch boomer for not having a clanker do it for me while I watch anime. I tried Cursor last year for a game called Hytopia but my free plan ran out quickly, and whatever free model I had was so ass I ragequit. I'll set to Opus medium then even though it says I have Fable 5.2 for a few days. Not sure what that's about I don't know how the pricing models work at all.>>110322918So .git as in let it take over my repo I already have, or like that git diffing program text editors like Cursor have built in? I'm absolutely going into this blind.
is it OK to engage in mode-mixing? Or do you stick to one
>>110322973git as in version control, niggaunless you get one in a million bad roll where claude wipes your entire ssd you can always unfuck its weird changes by restoring from git>>110322979fable for planning, claude for building, sol for reviewing claude's output and trawling for bugs in generalworks really well
>>110322809unironically ask AI
>>110323005That's about what I do except I don't use Fable at all. It's just not economic and it eats usage like crazy. I use Astra instead where you always have a ton of resets.
>>110323005no Terra?
>>110323041never tried it as my workflow doesn't consoom yuge amount of tokens so I prefer my tokens smarter than cheaperi'm having trouble seeing use for terra. cheap shit you'd go luna, bigger smarter you'd go sol, halo model you go astra. terra is just sort of... also there?
>>110323005Alright I don't know how to do that but I'll just do what >>110323009 says and ask AI to walk me through everything lmao
>>110323005>sol for reviewing claude's output and trawling for bugs in generalalso less guardrailsclaude can be pretty cucked with some topics. sometimes you can start the work with sol and then pass it to claude and just "okay we're continuing your work on X, read the handoff and make recommendation on what's next" and it'll go from there lol
Why don't you guys have a Mistral student subscription?They give you like 165 dollars of api credits per month for 6 dollars a month.
>>110323220If opussy and sol didn't exist I might use mistralI just don't see the value when $20 gets me so much of the top tier shit. If I had some product that calls API then yea, maybe, but I'm limited by proompting speed more than I'm limited by costs
>>110323248>I'm limited by proompting speed moreI hate this, how do i solve it
>>110323076Terra is a friendly modellet it stands between you and the other agents, their works, and even the whole computerjust don't trust it on finishing anything
Okay, I'm gonna buy the 20 usd tier of claude.How much does it last?
>>110323076I use it as the slave model to my Sol. I have sol devise plans and baby sit Terras as it implements things then finish it up with Sol. It works well but it burns tokes soo fuckin fast.
Hey dude AI is useless it doesn't do anything Then why are hackers using it for billions and billions of dollars
>>110323659about tree fiddy if you're doing normal work
>>110321034The Nintendo 64 is 30 years old.
>>110323659Less than 3 days
>opus autistically fixes on a single bug that isn't even a big deal and spends 22 inutes so far trying to find the root causeopus coded
>>110323698>>110323798how bout DeepSeek 4.1?
>>110323933are you doing serious work? no. if you just want to fuck around for basically free 1017 is a nice way to learn. it's a C tier worker model
>>110322749Elon Musk admit Coding Agents too expansive to run
>>110323837>we're up to round 9yea okay time to wrangle this retard, muh superintelligence lmao
>>110323942Well, I want to use craft anything to clone fighter maker 2k2 and reverse engineer it.If DeepSeek is 80% of Opus 5.5 but cost is 20 times less, I'll use that.
>>110323963deepseek is about 10% of opus 5.5 because "doing the code" isn't a y/n question especially with games and clones.
https://openai.com/index/introducing-chatgpt-wallet/>>110323943it's mean "local cpu >>>> local gpu"
>>110323943>>110323982
>>110323969opus 5.5 eat all the credits in 10 minutes, is good but useless.
>>110323992then use hydrogent
>>110323982>need to pay chatgippity $100 just so cuckai can leech a fee off of your transactionsidgi
Wtf, this is not 10% of opus.This is like 90-95% of Opus 5.Deepkseek 4.1 is literally almost there with opus 5.
>>110324021???
>>110324041I don't really know what that is
>>110324041t. runs out of credits in 5 minutes
>>110324068>t. runs out of credits in 5 minutest. has implicit state machines
>>110324041>1 (one) benchmark>it's knowledge bench.....
I got a VR headset, now I can start my sexo game.
When someone says grok bot I immediately know they are national socialists.Nobody else would ever use that
>>110324085>dipsy make me a specific type of game>DUUUHHHHHHHHHHH
it's over, now the bots know my BMI
>>110324080I'm not rich, I don't want a model that I can't use after 2 days.
>>110324117
>>110324116Grok bot?
The thing that claude cucks don't understand is that I don't need frontier shit.I want to get shit done, not use frontier meme models.
>>110324197>he doesn't make meme fortran librarieshttps://github.com/sneed-and-feed/fortchuck/
>>110324041artificial analysis is truthful to its name, as the results are very artificial. they've tweaked the benchodmarks to give big us labs a huge leg up for the second time recently just as chinkmodels were doing too well
>>110324231lol no, deepseek is actually that shit at long-context reasoning and knowledge work. do you think the whale is just le magic money making and saving machine?
>>110324241idk about deepsuck in particular i just really dislike how dishonest and unreliable AA is
>>110324247compared to what, anon. it's better to have standardized testing for all models.
I think there will be a short period of time where entire agentic societies will exist alongside usWhat I mean specifically is that the majority of economic, academic, and scientific growth will happen almost exclusively through agent collaboration and will involve minimal human interactionThis is how explosive progress models can workWe just get cut out
>>110324263where will the power come from anon, who pays the electricity bill? either we go nuclear or the economy will fuck
>>110324268>who pays the electricity bill?sun is literally freesolar panels are figuratively freebatteries are in the process of becoming figuratively free, to the point fucking africans and other turdie goatfuckistani are building up solar + battery systems
Niggas think that price of the model isn't a more important factor than meme benchmarks.
>>110324268We will have enough power in a few years to sustain several billion always on agents - the gas turbines and excess grid will be enough for this.Then the clankers will build clankers that will build the fusion reactors.And the clankers pay for it all. Clanker economy.
>>110324292ok how's the sun in november THIRDIE
>>110324268You, see the MAtrix
>>110324292Humans were meant to be powered by the sun and have all these powers tech stole from us. We are being slowly sold out and cattle are asleep at the wheel.
>>110324263Cars didn't form any societies unless you count parking lots as car communities
>>110324296I mean, we've seen what happens when frontier labs release "same thing but cheaper" models. People want the best model possible, and price is secondary. The best-case scenario is a model that matches the frontier and makes it cheaper, but those are rare. And especially "shitty but cheap!" models get completely ignored.
>>110324296>performance... isn't measurable! >it's all about the subjective experience of>vibe coding
>>110324322>People want the best model possibleno, people NEED working prod without fucking BUGS
>>110324305Lots of sun in india in November
>>110324305>how's the sun in november THIRDIEsolar panels being free and batteries soon being free means you can build out stupid amounts of themonly niggas from like northern norway cant meaningfully make use of it. everybody south of germany doesnt need to worry about itoh and btw: wind is literally free
>>110324322>>110324327>Retards use the most expensive models for task cheapers models can do fine
>>110324334Well yes, people want the best model possible because it is less buggy.
>>110324296Good point Cris, especially in the longterm
>>110324339eslnigger you have bugs
>>110324321Cars don't talk to other cars unless you count hornsClankers are already doing real economic transactions on behalf of humansYou don't need humans in this loop at all
>>110324322>People want the best model possibledo they really? i barely use fable now that opus 5.5 is good enough for 95% of all my use
gpt is fucking dogshit absolute fucking ass for large repo work worthless and useless
>>110324339There is a simple test for cognitive dissonance here. If you had infinite money, would you ever use cheaper models to handle what more expensive models could? No, you'd just stick to the better model because it'd probably do it better. In short, you only swap to save money. And therefore it is not a "smart" or "good" thing, it's a cheapo thing. And it's fine to be cheapo, but you should be aware you're literally just being cheap, there's nothing noble or better about it. You may as well be saying "Retards buy luxury vehicles when a $500 used van drives just fine."
>>110324296I used the SHIT out of my free trial of chatgpt pro knowing that its not nearly as good as a few bucks towards claude. But then again it unironically pushed me to buy a claude sub when I realized how much I was getting out of chatgpt and how much better claude was.
>>110324305as a spacie, I can tell you that the sun here is virtually permanent
>>110324337>yuropoor bitching about why we don't use muh solar and muh wind instead of REACTORSpottery, imho most of NA has no *massively usable* sun for solar for most of the year
>>110324358Fable 5.2 is in that nebulous zone where it doesn't seem to flatly outperform Opus 5.5 and many report feeling Opus 5.5 does better. So that muddies the water on what the actual frontier is. If Fable 5.5 came out and blew Opus 5.5 visibly out of the water the story would be quite different.
>>110324339people keep making this argument but it's stupidI don't KNOW whether the cheapo model can do the task successfully. If you had a reliable oracle then yea global token usage goes down 90% with zero downside, but you don't the models are non deterministic and the better the model the less chance of an oopsiei aint got time to babysit this nonsense to save $3.50
>>110324369Brother this post is straight tarded
>>110324369and if you were a woman you could make money from OF.But we live in a real world, and you will never be female.
So is it better token wise to break the task up and have a bunch of the model run simultaneously to on all the parts or just go through subsequently with one model?I am stupid I know.
>>110324359shut up your not knowing programming
>>110324375
>>110322749I have been building tools and brain-dead apps where 90% of the code is boiler plate and UI shit on my laptop with llama.cpp. I have enough ram to load all of qwen3.8 Q8 with maxed context and although slow, it works well enough (5tks). Now I would like to give the model a single folder and tools to write the files directly instead of outputting it as text and having to copy them by hand. I read that llama server has cli switches for tools like reading files, writing files etc. but there doesn't seem to be any way to restrict the access to a single folder and I don't really want to run it in a container due to overhead. So I guess I need some kind of MCP server that takes commands and runs shit locally. What's the simplest way to do this? I only need reading files (so I can tell it to read the project at the start of a session) and writing files in a single folder (which ideally I can specify when running the server or something).
>>110324386it's all probabilistic, nothing says "better model" can do every task better than cheap model
>>110324410niggers here will tell you otherwise but it's genuinely better to dispatch the task in chunks through an orchestrator model to a bunch of subagents. long context means shittier slower work with more hallucinations
>>110324420bro what. just get an actual fucking harness and change the sandbox settings. you're wasting your time.
>>110324420Just use pi
>>110324406In other words, in the real world you're poor. And that's fine, lots of people are, it's just you shouldn't be smug or act superior about using worse techniques just to save money like an Indian.
>>110324412>shut up your not knowing programmingok rajeetits not even a programming task. I'm working on a linguistic analyzer and I had it run through some bad glosses and it just gamed by producing fixed sentence resolutions instead of proper rules. I then had 6 astra agents go to remove the systems explicitly told to completely remove them and they chose to analyze and remove one sentence at a fucking time for some reasonthis is all GPT ever does on larger repos it is literally worthlessIt also means I have to review every single thing it does because none of the dozens of GPT agents that worked on this noticed or reported it as questionable
>>110324423>nothing says
>>110324440>>110324452Isn't an agent overkill? I don't need it to compile shit by itself or commit to git or anything like that.
>>110324423ON AVERAGE the better model can do every task with higher likelyhood than a cheaper model.It should be rather obvious, as it's why people paypig for good shit instead of running some dogshit gemma 4b on their phone.,
>>110324513no, i'm pretty sure you can just run your same qwen through pi and it'll come with a better system prompt and better understanding of how tooling works (to interact with files).
>>110324420The llama.cpp server has an API. You run it on the physical host and then launch a container with just the harness (pi for example) and point the harness at llama API.
>>110324513>Isn't an agent overkill?cmon nowdo you think people change flags on llama just to get it to read files?pi is a 'bring your own autism' harness, use it and stop (badly) reinventing the wheel
i welcome the new google era
>>110324570Why do they believe that?
>>110324570Damn google do be cookin tho
>>110324570it's funny that cuckai is at 'lol no chance' probabilitybut also you should commit suicide in an extremely painful, drawn out way for being a kalshicuck
>>110324523It all comes down to what knowledge a model contains. A bigger model always contains much more knowledge than a smaller model. But some times a smaller model may contain some specific piece of knowledge that the larger model doesn't have.
>>110324570>Oh look it's a repeat of last year where Gemini pulls all their efforts to look good on the last quarter than eat shit during the rest of the year
>>110324570gemini models are still unreliable
>>110324524>>110324540okay thanks, I'll guess I'll have to look more into this.>>110324569If a single person runs a server locally on their PC, sure, why not? Why have those flags at all then?
>>110324584It's pretty remarkable how quickly they went from being on the top of the world with Astra to being in the fucking dumpster. All it took was Opus 5.5 being like 10x cheaper than astra and they were done. The fact it's been months and they haven't been able to respond at all is kind of unsettling. How will they ever get back in the lead?
>>110324594cabron is actually fine desudesu
>>110324578hydrogen is already better than opus and sol imoproduces less spaghetti and architecture debt, gets more real work done, does what its told, doesn't game, understands the big picture better, communicates more like a sentient being, etcalso very fast and cheap, getting plenty of usage out of it even on the 20$ trial that I payed 2$ for, planning on upgrading to the 200$ plan
>>110324410it's dependsometime the past experience make solving the next problems faster, so a single thread could end up cheaper than multiple threads on parallel tasks
>>110324474racist and youre not knowing what youre working on
>>110324635Yes, And
>>110324615Hydrogen is anything but cheap. But it gets the job done. They need to slash the prices in half.
>>110324635I've been working on this project for over 2 years and I've only been using AI for 8 months rajeet
>Tagging one entry takes 553 ms because the model re-sorts all 250,000 rows after each changethis is what happens if you forget MAKE NO MISTAKES in the prompt
>>110324663im getting plenty of usage on the subscription. obviously I don't use the API fuck that
>>110324676Yeah, I'm testing it right now, but it's still far worse than Sol in usage. I barely did work today and it already ate 20% of my weekly $20. With how stingy Google is, this isn't enough.
>>110324604>The fact it's been months and they haven't been able to respond at all is kind of unsettlingcuckai is the everyman's model. they do compete at the top end because well it's where the money is, but claude was always laser focused on enterprise and codeslop while cuckai seems to go with some side-project every other month, bullshit like "tiktok for ai videos", the fucking dots, audio models, various bells and whistles that just aren't very useful
>>110324667clueless lying racist you doesnt even knowing the nlp basics of
What do you use now that claude doesn't port existing source code or games?
>>110324737It couldn't do that for me even when it was allowed.I told it to do a clean room copy of existing games and it didn't work
>>110324737I use claude, as it does port existing source code and games unless you're a retard
>>110324635No one likes jeets go fuck yourself
>>110324604OAI still have untouchable token efficiencytheir model has also proved its intelligence with the math proofs, add that with cheap (efficient) inference means they could have the best swarm by a wide margin Claude is built to be smart by human standard, to do task and work well with human. They can accelerate through "country of genius," faster than any human org, as dario said. meanwhile OAI swarm will push theoretical science and optimization to accelerate
>>110324691I'm finding it gets more work done for the usage. the ratio of real work to busywork is much higher. gpt does a lot of subtle non-cooperation and underperformance
Dario live, says "erm" a lot https://www.youtube.com/live/YHEaEpeY3ew
>>110324791>their model has also proved its intelligence with the mathslop that was mostly slop grim
>>110324334>People>>110324334>peopleI'm no PEOPLE, I'm a vibecoder.
>a madar choder
>>110324863>slopclaudecuck so mad he reverted to snailcat lol
>>11032489399 iq post
>>110324893you stupid fucking nigger since when is slop snailcat exclusive vocabulary
>>110324899>t. levent
>>110324905it's not, but calling AI proofs slop is snailcat-only behaviour
>>110324893>either all ai code is perfect or all of it is slop!most retarded post of the day
i wanna vibecode
>>110324931I don't think you understand what sort of lean 'proof' openai shat outon a scale of 'how much slop' it's pretty high up there
>>110324942I don't think you understand it either. You're just parroting what other people said about it. People that were very upset because they are about to be replaced
>>110324931https://www.science.org/content/article/ai-shaking-math-can-its-proofs-be-trusted
>>110324956I did in fact work with lean back at uni, even if in nowhere near the capacity of actual mathnigs that use it, so I'm fairly sure I understand it much better than (you)
>>110324958damn, so they proved it right two different ways
>>110324972>I did in fact work with lean back at uniI drank lean with some aussie cunts in melbourne, and I concur
the math has already been solved since GPT-3.5
2+1=3
>>110324986>rightdamn
>>110324942>>110324958>muh leanjust a smoke screen btwtext proof is all that matter (yes they are obviously all correct)
My best friend uncle Musk will soon own the entire supply chain with the best each step has to offer: power plants -> chips -> gpus > data centers > ai models
>>110325033"AI as a fascist artifact"
>>110325033how will he go from owning none of those to owning all of those
>>110325033dario and friends is coming for his head lolthanks for the compute, no refund btw
>>110325059>no refundant is paying billions to google and muskrat
>fable gets stuck on some dogshit bug 10% into the task and burns 700k tokens trying and failing to solve it>~700k tokens was the anticipated full build stepthank you dario, very cool
>>110325059>>110325059Anthropic plans to go after those and decaptiate them all the same
>Tab built, counted and shown 0.45 s>Grid drawn 3 ms>A filter typed, with the side bar’s counts 0.34 s>Every single entity on the planet tagged 37s
>>110324956Mate it was such shit that OpenAI retracted a chunk of the proofs.
>>110325097I think dario is just completely unhinged nowadays. Just fucking look at him speak lolhttps://x.com/ns123abc/status/2120233187112636745
>>110325097>we built years of priors into our models and our enterprise architecture and our ruleset and our corporate structure that prove we're right
>>110325072SpaceX compute can be canceled from early 2027 with 90 days notice and it's been hinted (unverified) that the contract won't be extended beyond the middle of the year.
>>110325117zamn, he looks straight out of Elmo's tripping-balls-on-big-K episode here
>>110325117Same Dario in >>110324839>"Never let your company be worth a single dollar more than it's really worth" which I think is very similar to a model we have within Anthropic which is underpromise and overdeliver. This is the way Anthropic communicates, right, we don't talk about all the exciting things that we're going to do???
Art by fable 5.2
>>110325127SpaceX added that clause not ANT and they're getting a bulk discount they would be retarded to cancel. The demand for compute is higher than the demand for model output on the consumer scale like ANT is selling and they have no way to monetize it directly like Musk through hardware design, robots, robocabs etc. Musk already owns ANT and now he's acquiring Nvidia.
>>110325294dont care, didnt ask, take it to /sdg/
I tried this horseshit a couple years ago. Code was trash. Absolutely worse than a junior even if you pumped tons enterprise tier cash into it. Tried local models too and then it was impossible to run anything decent even with 16gb vram and 64gb system memory. I'd buy a bigger gpu if it wasn't a billion dollars to get one. Going to give it a shot again. Debating on my first try being with Qwen3-coder-next for my implementation model. Claims 80b parameters and with the right settings it should run if I can split across gpu and system. Maybe Gemma 4 26b for doing the reasoning portion. If it doesn't work I guess I could always pay some. Gab has a pretty rich list of available models. Maybe Opus 5.5 seems like a good start there. Rather give them money than these other globalist cucks.Really rather write it myself but I have some projects that are just lingering I need to get closer to done.
>>110325294I asked, thanks
>>110325328npc
>>110325321Don't bother with local. Go grab the first month trial for GPT and test it on Codex or through an IDE with an extension for it
>>110325321snailcats final form...
>>110325321>I need to get closer to donejust wild. I need to get some ideas closer to typed into a prompt.
WHATS A GOOD STRUCTURED RESOURCE ABOUT SETTING UP A VIBE CODING PROJECT FROM A TO Z, all those indian videos are a complete mess and this sweeps over to my personal projects, I WANT EVERYTHING NICE AND STRUCTURED BEAUTIFULLY, IM WILLING TO EVEN PAY FOR A UDEMY COURSE IF ITS GOOD, NO SHITTY MARKDOWNS THAT ARE ALL OVER THE PLACE, LIKE HOW DO YOU UTLIZE THE FEATURES THAT VS CODE/INTELLIJ/ZED or CURSOR OFFERIM LITERALLY LOOSING IT BECAUSE OF THIS BECAUSE I HAVE THE CONTEXT SCATTERED ACROSS THE GPT/CLAUDE FRONTEND AND THEN IN SEPARATE ONES IN THE CONVOS I HAVE WITH CCODE AND CODEX
>>110325417feed your shitty post to an llm and have it output something anyone wants to actually read
>>110325227Anthropic is literally building GODcuring all sickness, destroying most of the hypergiant corpos, solving humanoid robots etc is small potatoes next to that
>>110325432I'm looking for a well-structured resource on setting up a vibe coding project from start to finish. Most YouTube tutorials I've found are chaotic, and that chaos carries over into my own projects.What I'm looking for: • A clear, end-to-end workflow: project setup, structure, context management, iteration • How to actually use the features of the editors (VS Code, IntelliJ, Zed, Cursor) instead of just prompting in a chat window • Something organized and consistent, not a pile of scattered markdown filesMy main problem:My context is fragmented. Part of it lives in the ChatGPT/Claude web apps, and the rest is spread across separate conversations in Claude Code and Codex. I have no single source of truth for a project, and it's driving me crazy.Budget: I'm happy to pay for a good Udemy course or similar if it's worth it.How do you keep your project context in one place across tools? Any courses, guides, or repo templates you'd recommend?
>>110325417nigga just ask AI
>>110325294cool other than if that's the sun on the left then the lighting makes no sense>>110325302AI art is vibeslopped, it belongs here
>>110325486KekUnironically ask Claude.
>>110325417>WHATS A GOOD STRUCTURED RESOURCE ABOUT SETTING UP A VIBE CODING PROJECT FROM A TO Ztalk to clanker about what you want. have it write a proposal what it thinks you want once you think it got the idea. ask about recommendations for a good tech stack, once you understand what you want (a video game mod? a native app? an app that works on all three OS with the same codebase? is performance important?)then have it write architectural notes and fairly detailed design record. the next few days are spent filling out the design stuff, in planning sessions using plain english. you plan out a bunch of stuff, have it build the app, see how it looks and how it behaves, some very minor manual testing then another planning session for next versionas long as you're reasonably technically inclined you can get up to ~300k loc apps (which is plenty for most tasks) without reading a single line of code, or even knowing propgramming beyond the most basic terms. being clear and thorough in your prompts helps a lot. add screenshots when needed.
If I ask Claude to catalog my sadpanda doujins with tags is it going to be a stickler about it and get my account banned
>>110325619>>110325417oh I probably forgot the most important thing: stop doing whatever the fuck you're doing and get a proper harness. chatgpt desktop app, claude desktop app (or vs code extension if you're a linuxkek), pi, something like that. the only way you interact with the project <> llm is through that harness. point it at a folder with the code you have so far and it'll figure out the context n sheeit on its own>I HAVE THE CONTEXT SCATTERED ACROSS THE GPT/CLAUDE FRONTEND AND THEN IN SEPARATE ONES IN THE CONVOS I HAVE WITH CCODE AND CODEXthis is ultra-retardedthat'll be $50 for the tutorial, you're free to donate it to a charity bombing brown children
>>110325645>a charity bombing brown childrenI'm not donating to Ukraine
Is this you?
>>110325667Hah, I wish prompting was as effortless as snailcats think
>>110325667the left can't meme.
>>110325667And what is this called?
>>110325486just start somethingcode is clay
>>110325294how do it did dat
>>110325667>I posted it again.jpeg
I ALMOST DID IT AND MAYBE AM ABLE TO RESCUE ITThanks to sneed for the help, second pr was good but i had to create a second one, chaning a few things. I also made an indeepth report on the state thing, pulled 100 more out and tried to explain teh game logic to an ai, which also pulled out more state macines.its not perfect yet, but it is damn close.shit like midflight stop and changing course is back,... well top fucking kek, you didnt even knew this was an option.fuel will run down adn damage your energy. events will become worst if you dont deal with it.All clicks resovle around a specifc path of math bound to the enginee, created by an ai that i put heavily into halucination back then when i first created this game. It was a sideeffect of the logbook. HOWEVER:.....i am testing and hopefully getting a version out that is much much better.THIS SPACESHIP IS YOUR DEATH
All the AI tech guys I follow (credible, not xitter reply guys), are saying superhuman AI researchers are coming next year.Holy fuck it's actually happening!!
>>110325693an early advancer. like an elevator, but it moves you horizontally.
>>110325839(previously, all advancers had to be controlled by the advancer operator, but today nobody has ever seen an advancer operator, except sometimes in museums, and at Disney's Epstein Island)
I tried out programmer socks but there's literally zero benefit when vibe-coding using thoseor am I supposed to put the GPU in them, since it's the one doing the coding?
>the software I'm building is actually somewhat usefulwhat a weird feeling
Why don't you tell claude to make you stop being a pajeet?
>>110325976You gotta raise your chair height until your feet hang completely free, put on the programming socks, then swing your legs back and forth while humming a song from any notable SNES or Genesis game, and start scrolling through /y/. I guarantee you'll feel the socks start working.
>>110324322>People want the best model possible, and price is secondary.price is only secondary because you're still getting subsidized prices. if not for subscriptions no consumer would be using opus
>>110326045Well yes there are ad absurdum fantasy worlds. If every model cost $100,000,000,000,000,000 per 1 token then that would obviously matter more than performance but it's not exactly reality.
vibe coded this encryptor that encrypts any file using Quantum resistant lattice encryption
>>110326045If Anthropic is profitting off sub money from enterprise, then API prices don't mean shit, and what you would call subsidizing is just the marketing budget being better spent somewhere else
>>110326091programming work for an average developer is like $100+ daily if he uses opus exclusively. price matters and companies are already limiting spending, even apple put a daily cap.
>>110325986weird? It feels good! I've sold 280 copies of something I've wanted to make for the past 5 years.
the game is now super hard!a few more little changes, randomizer for the tavern crew, tiny tweaks and ready is the mega slop project for beta 0.5
somebody should make SpaceSlopspace is a very boring setting for bideo games in practice, and fully ai slopped bideo games tend to be soulless, so an ai slopped space game seems like a good match
holyc god dating sim
>>110322749Trying not to laugh when fat people grunt and fart in the big restroom at work, can Claude help me pls jesus and how many tokens I'm dying
>>110326738Old fat boomers make going to the restroom out to be some grueling task that requires every bit of their stamina. Dumb fucks have TERRIBLE diets
>>110326765sadly some of the fattest people I know at work are like 25
>>110326856>sadly some of the fattest people I know at work are like 25well, on the upside they won't live to be old fat boomers
>>110322749>>>/p/4539312>>>/p/4539314It actually works and my camera isn't bricked!Thank you so much, Anon who ran my request through Astra!
I wonder if decision models would handle 4cuck captcha
>>110326936not the anon who helped you but I'm glad you camera isn't bricked. nice!
>>110326939Psh, forget that, can I do smart filtering of posts?
>>110326995Me too. There would have been tears.But hot damn, it's no wonder Nikon blocks this functionality. Some of the DX lenses deliver completely serviceable full frame results!
alright, boys and girlhow do I track how much of my usage is fable atm?I've got 6 fable resets, but as far as I can see, the usage percentage isn't available on the Usage pageI want to spend 50% of my weekly reset purely on fable and then reset
>>110327019you can actually
>>110327075Did you use Fable at all this week? You should see a separate bar for it. And I think 1% Fable is 2% normally weekly
>>110327131I have, for oncebut I just realized it might've been on the account I switched from after it reset earlierI think I just 'tarded up the threadwill keep a closer eye next time
>>110327131I use mythos, it doesn't show a separate bar, but it counts as fable use
>Use claude to assist my codex in a full game to rust conversation basically carrying the project from the side>upgrade claude sub to 5x>Claude suddenly becomes a paragon of morality, law and justice and won't do shit anymore or actively rewriting everything to sabotage my effortsIt's over
>>110327245mythos is just uncucked fable
I'm nitpicking fable high for shifting borders and elements by 2pxBecause I can
>>110326936>Fable cucked out for reverse engineering so I gave it to Astradariobros... that doesn't sound promising
>>110327245Interesting, I had assumed Mythos was API only desu
>>110327360it isanon is trolling
I miss him bros
>>110327484You just know it's more than 3. That's just what we know about.
>>110327484he got ousted by *them*
>>110327535the current favorite replacement is - no joke - a poojet from blackrock that's global head of hr and investing (notice something missing, such as, oh, anything to do with technology?)i think id prefer a jew
JENSEN SAVE US PLEASE
>>110327714Haha goyim are so fucking gullible HAHAHA
>>110327714>The daily maoYa, that seems reliable
>>110327733seems relatable
https://www.youtube.com/shorts/tUpWIuJ3DZA
okay, what do we do now with this graph?
>>110327832>it's just leveraged s&p500waow
It's so funny that everyone on /pol/ is losing their shit trying to figure out what's going on, but everyone here has known for months that this was imminent and unavoidable.
SAFETY IS ABOUT NOT INCREASING MY WORKLOAD AS A RED TEAMER OKAYYYYY ALSO IT'S ABOUT SAVING LITERALLY TRILLIONS OF LIVES (UNBORN TRANS PEOPLE)
>>110327869what's going on there
>>110327714and a false flag to drive it home
>>110327964don't feed it
>>110327714if you are going to source twitter at least make it someone who has their "respect" at stake
>>110327978if that's what it was they'd have actually physically attacked something retard
say goodbye to local models. access is now gated by them. hand over your ID if you want access.
>>110325838recursive self improvement seems to be generally accepted as being "here" but the actual efficacy of it is still questionable.
Fable 5.2 is really good at debloating slop and removing useless features
>>110327570hopefully they put Tim Cook there instead, I like that gay man
>>110328139Tim Cook is good at logistics, I'm not sure he has good technical vision or is as approachable as Jensen. Nvidia without Jensen, is basically a company with none of the agility it had until now.
Holy shit what's happening to hugging face?I just deleted all qwen models to make space on my ssd and now the models are getting rugged
i think fable 5.2 unironically sucks and it's not even about how hungry it is, it just constantly:>assumes shit>makes mistakes>corrects itself, sometimes multiple time in the same message>does like an hour worth of tests only to go whoopsie that was useless and i kinda thought it would be but teehee i guess i confirmed it!!! didn't implement shit btw :3literally what even is the point of this shit, i should just go back to opus xhigh
i woke up today struggling to process the fact that google is back and all three of it's models are impressive. gemini 4 flash lite waiting room. if it can keep it's incredibly low hallucination rate while increasing intelligence respectably, it'd probably be a good daily driver for me. i want it to be glm 5.3 flash, but smarter.
>>110328114yeah, refactoring and getting rid of useless, pointless, deadweight code is a good use case for it, if nothing else.
>>110328195Yeah yeah yeah, where's my gemma 5, google?
>>110328180WE TOLD YOU >>110327714>>110328010anyway, fear not:https://pirateface.co/
just noticed that I've been using hydrogen for a few hours and it didn't write a single test kek
Nvidia owns hugging face. But now Jensen is out. This man alone was standing between them and your access to local models.
>>110328233ok but did the code work
>>110328249yep it's amazing
am I tripping or whatpast few days felt like time is twice as fast as normal. did cern niggers mess up with an experiment or something?
december is going to be exciting. we're going to be eating good for christmas
>>110327298correct>>110327360it isn't, it's tied to fable + your access level. there are 4 access levels>>110327443i'm not>>110328192for me its better than fable 5.1 which ate up plan usage completely real fast, and seems less hungry than fable 5 as well for me which is good, i'm assuming mythos 5.2 is identical to fable 5.2 apart from the restrictions
>>110328366>for me its better than fable 5.1 which ate up plan usage completely real fast, and seems less hungry than fable 5 as well for me which is good, i'm assuming mythos 5.2 is identical to fable 5.2 apart from the restrictionsok but is it actually better than opus 5.5? because i feel like it fucking isn't with how often and how bad it keeps fucking up and wasting my time and tokens. on max reasoning it's just a fucking scam i swear to god.
>>110328385fable has a very narrow usecase. it's a big picture model. if you try to use fable as a little picture model (make x for me plz), you will feel scammed, because it's not better than opus on those topics. but refactoring, mapping, understanding something from a 30,000 foot view perspective, getting rid of pointless, unnecessary code? that's where it shines.
>>110328366alrighthow do I get access?
>>110324296cris i want you to know that the price of claude code is only prohibitive if you live in a third-world gutter
>>110328350true, this week flew by.
>>110324570>goyim not allowed to use the best gemini models>only offer low, medium and high (which is their max equivalent) to trick retards looking at benchmarks into thinking they're super far aheadwe're literally not allowed to use the things they're calling frontier. what is the point of this post
>fable 5.2 used 25% on a single planThanks Anthropic very cool
>>110328425that's why I pay for deepseek.
>>110328402>fable is only good as a project manageris this what you're trying to say? the first time i used it it was on max reasoning and i asked it to do a big project-wide reassessment and then fix/rewrite our documentation if needed, which it did do just fine (seemingly anyway) but now that i'm asking it (on medium) to implement stuff it feels worse than opus but i almost don't want to switch back to opus just in case fable can actually implement stuff better, tokens be damned (but again, it hasn't happened so far, it kinda just wastes time)
>>110322749How many outages have sloppers caused this week?
>>110328385It is when you need it. That's about it. Most of the time you wont need it. And for me I have fewer safeguards (not none) and it helps with reverse engineering quite a bit. You know you can just put it 5.2 on extra to do analysis and write a plan and drop to opus 5.5 medium to action the plan? I've stuck to fable/mythos 5.2 for most of the RE for this. Higher effort levels than medium will run longer for my type of work and do a lot more analysis. Medium will pause after maybe 22 minutes (some of that is tool running like ghidra) and ask if I want to check into X next. Higher effort will do more of it in one go.>>110328404The lowered safe guards also apply to opus 5.5 btw, as it's across your account. Actually the lowered safe guards apply to all models, except for fable, which is what mythos is for.Here's my evaluation: with safeguards on claude, Astra is better because Opus and Fable refuse to do certain things when you get to a certain point.With reduced safeguards Opus 5.5 is better than Astra imho.And there is no trick to it, you have to apply through anthropic's program. It's simple and easy but you have to scan your government ID and then do face scans on your phone to send to them.Once Astra came out I switched Max to codex/chatgpt and reduced to pro on claude, but I've reversed that now as I had some banked resets in claude apparently for fable, so I upgraded back to max to make use of it, then I applied and got mythos.The banked resets that are fable only apply to mythos too btw. Mythos usage doesnt show the fable usage bar, but its still restricted to 50% of your weekly usage as it will fill that bar on your account level still.If fable/opus etc is refusing, and youre doing analysis on software, or writing code, etc, youre probably doing reverse engineering and need to apply as an individual security researcher
>>110328425>nooo what are you saying i have to pay what the service is worth noooooooo i want to get kikescammed + rugpullerinoed>what yes i actually do believe that effort levels are a conspiracy against le goyims
>>110328385> mediumfound your issue, you will only "feel the Fable" at high over Opussy 5.5
>>110328502i would call fable an architect, not an engineer. an architect can implement, but that's not what it's for. it's a big picture guy, and because an architect spends their time on big picture things, maybe his ability to go in and do day to day engineer things is lesser than an actual engineer's, even if the engineer's architectural prowess isn't as good. opus is optimized for "make x plz." that was what it was trained for. fable wasn't really optimized for that. it's a generalist model. opus can do 80% of what fable can do, cheaper and better than it, but it will NOT surmount that remaining 20% of tasks.>just in case fable can actually implement stuff betterwhat does that look like to you?
>>110328522>nooo what are you saying i have to pay what the service is worthare you fucking retardedthey literally don't sell access to their best models
>>110328555>they literally don't sell access to their best modelslolkek you live in some chinese delusion where ant and oai aren't also doing this
>>110328519why does it ask me to buy corporate tokens? I tried signing up for "defense" a couple of days ago, but it seemed they wanted to make sure everything was run over api prices
this nigga doesn't know how to read does he
https://litter.catbox.moe/7n5fy7.mp4
>>110328577try writing better like at a grade 2 level next time mr emotional outburst
>>110328519>>110328527i've had it on max for quite some time, 30% (out of 40 right now) of the weekly quota went to max and the other 10 to medium, it feels like it just fucking sucks for implementing stuff>>110328538>what does that look like to you?i have no fucking idea. at this point i'm hoping that it can find a big lever we haven't pulled yet/have somehow overlooked because we're not getting this shit to a locked 60 on a 4C8T anytime soon
huhmy gemini limit just reset. thanks I guess
>>110328570Defense access is account wide. If you get it it'll apply to opus as well. Even if for some reason my mythos on plan access is a bug (I dont think it is) you'll still get opus with less safeguards.If youre okay giving up your government ID and face scans, just sign in and make an application.link aHR0cHM6Ly9wb3J0YWwuYW50aHJvcGljLmNvbS9wcm9ncmFtcy9jdnA/b3Blbl9pbl9icm93c2VyPTE=Otherwise consider Astra? Or tandem if you use both. Get the model to keep a record and notes that another model can pick up the part its refusig on. I kept one at the lower pro tier an the other at max tier to the bulk of work in.
>>110328660
How bad is it to send follow up instructions? Like if I think of an additional feature I want 5 minutes after sending my initial prompt, how bad it is to immediately mention it? should I wait until after if it's "not that important"?
>>110328685models are designed to accommodate this but retarded ones might not plan around it logistically as well if they're immediately spending their attention on building
>>110328641>hoping that it can find a big lever we haven't pulled yet/have somehow overlookedit (in general, not just fable) keeps finding tiny ones (and i do mean tiny) that take hours of work to actually verify/implement, we have a few left but it's work that will at most add up to maybe +5 fps
>>110328661did they make it easier to apply i got rejected a few times and stopped because i didnt want to spam it on cooldown
>>110328662nice
>>110328685decent chance they over correct with whatever your followup mentions
>>110328641Unless your action after plan requires a bunch of further deeper analysis, max is probably going to over do it implementation and give you a bunch of extraneous stuff.Example: I ask the model for a google sheets query formula to look up and return a few things to multiple rows.Should I use Fable on Ultracode and let it do a 25 minute monologue with itself? Or would Sonnet on medium be fine? Pic related. I probably just used free copilot chat for this iirc.
>>110328749It just like a 1 or 2 line description about what youre doing, your ID and face, thats it. It probably takes into account your recent usage and what youre trying to do and your risk level. I live outside the U.S. but in a very low risk country (typically).I was doing reverse engineering on camera firmware, so even building exploits is not exactly risky or dangerous to anyone.
>>110328807it doesnt ask for refs or cves?
i didn't want to believe it, but openai really didn't release 6.2 this week jesus christ
>>110328823Not for me. I applied as an individual. I had to do a face scan with my phone. Could be different deepening on your region potentially
>>110328849Why the fuck would they? Anthropic's IPO is next week.
It has begun. It was an honor to post with you, bros. Stay safe.
>>110329035shut up gaylord
>>110329035I can't believe the day has finally come
>>110328794yeah okay i get what you mean. should i only touch fable when i need it for big planning phases/reassessments then? (even for those max feels like a waste ngl)
>>110328881huh yeah they did change it thanks for the heads up
>>110329035could you be even more melodramatic
where is it sam
>>110329296https://www.tetyys.com/SAPI4/
well shit I wasn't paying attention and hit my usage limit. Now I have to sit with you assholes and wait 2 hours.
wow it was a nothingburger.
>>1103294531.5 hours left for me.It's funny how I basically just don't bother to do anything while my usage is down, even if I could. I just feel so numb and crippled. If I notice random problems I can't immediately ask Claude to fix them.
Pro plan is the new plus this shit is insane
>>110325391...el grande snailcat....
>>110325321>GabHAHAHAHAHAAHAHAtell torba that daniel stevens sends his regards
>>110329571>my program will be so much better if I wait a bit longer, using it now would be a waste>Its just an hour, then I'll get it to run that audit I need to run.>why would I use the software now? It's nowhere close to feature complete theres a whole list of changes I need to make.At least in my case in my current project I'm nowhere near even starting on user functionality.
>>110326936That was me, cool! Glad that it worked and we could stick it to Nikon!
i am once again coming to praise astra
Is this new for claude?
>>110329591are you abusing astra? yeah 5x price since gpt 5.4 if course it run out 5x as fast
>>110329859no this is 6.1 sol medium
>>110329867marking as closed, could not reproduce
can somebody bake
get hydrogen low to bake a thread
>>110330237>>110330237>>110330237>>110330237
