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/----## 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>>110233352
what if snailcat became racistcould he compete then?
deepseek slopped up some fixes for me for my internet download manager app while opus was on cooldown. now that it's back, it's time for opus to check behind little dipsy. she was distilled from astra, so the code quality must be atrocious
Main character? Yes I am, how did you know?
>>110238390>how did you know?you seemed a little chromosome-heavy the moment i saw you
who would win?> Fable 5.5 vs Astra 6.2> Haiku 5.5 vs Luna 6.1
>>110238399fable 5.5.haiku 5.5. luna 6.1 is not going to be better than 6.1 terra, and 6.1 terra is marginally better than deepseek, which is slightly below haiku
>>110238399gemini falsh lite 5
ran out of tokens on my free vibe coding thingy. i don't want to pay a dime for tokens, how can i get an infinite amount of tokens for no money?
>>110238407you could do what aquaslopper does and run the arena.ai gauntlet on prod
>>110238206>>110238237Like I said, when you need more you'll know and you won't need to ask how to do it.Let me give you an example. I have been building an inference engine, I started with a handful of model integrations that I handheld through implementation to get structure, style etc., then I put the 800+ models I wanted into the backlog, this was a GitHub project, for each model in the backlog it listed the model's source, checkpoints, etc. I wrote a prompt that described the integration process, that it should write the model graph, compile it with profiling, run inference and compare to the reference output, and write a Rust crate that can use the compiled module with all the tokenizing or pre/postprocessing etc. in Rust, I was then able to just queue up, it would take batches of 4, and I was able to just queue up "Continue with the next batch", this of course took weeks considering the compilation, profiling and inference times. It worked fine, all the models eventually got integrated, passed reference validation, etc. However, upon inspection, I found that it had drifted from the initial requirements, there were a few different code styles, and various other issues, only the core part of a model/pipeline implemented like text-to-image not image-to-image for example. This is when I knew I needed something more. So I developed a custom harness, something using the codex sdk, this would again work from a backlog, this used a combination of prompts and enforced hard gates, meaning something that physically checks the code has followed the style, checks that there are no random helper functions, checks that it hasn't modified anything outside of the permitted scope, etc. if a hard check fails it gets sent back to the queue. I then ran that orchestrator with 8 workers, basically the equivalent of having 8 sessions but headless. I also ran an actual Codex session as a coordinator, this monitored the queue for stalls 1/?
nigga i aint reading all that shit
>>110238425and resolved shared things that were out of scope for the implementation workers.Initially the implementation workers would just create PRs, I would then manually review and add review comments on GitHub, then comment `/fix address review comments` or `/fix resolve merge conflicts`, this would get picked up by the orchestrator and send it back to the queue. Eventually I had done enough manual reviews to fold my new merge rubric into the implementation workers, then the implementation workers would follow things through to merge, while still leaving shared things to be resolved by the coordinator session.Overall this worked very well, after folding in the merge rubric I manually checked some of the merged PRs and found none of the issues that were enforced, in a very small amount I found issues like only eager attention instead of using flash attention because the original/source only had eagar, so I added that as a hard gate too. It's always an evolving process, you don't know what you need to enforce until you notice the issues.However, if your project isn't at the >1m LOC scale (mine is >2m), or you don't have parallel similar work, or whatever, then you don't need anything like this. If you're starting a fresh project from an idea then all you need is to prompt, no subagents that will only waste your tokens and time, literally just tell it to do the thing you want. My project started the same way, I just prompted what I wanted, then it evolved to using loops, until that wasn't enough and it evolved again, and so on until the point I'm at now.
>>110238433>>110238425Very interesting. What about the futanari hypothesis?
i think this wordslop was actually written by some schizo rather than ai generatedimpressive
>>110238463this is the guy burning billions of tokens to set up 800 rust crates lol
i just want dario sam and elon to apologize and kiss again and come back together and create SGI
>SGIoh shit a new gender dropped
I don't know why you keep making apps to do things when people can just ask AI to do things
artificial gender intelligence
>>110238472If you're too stupid to understand what I'm doing then just be quiet
>>1102384831) ai is adopted by less than 5% of society (at least in the us + europe + canada)2) people who use ai overwhelmingly use it for shit like writing emails, being a yes man who affirms them 100% of the time. they don't use it to replace a subscription app they have3) most people (myself included desu) don't know how to prompt ai, so it can be tedious to tell ai to fix problem after problem after problem, when the handcoded app is seamless and polished. people who know how to prompt don't have this problem.
>>110238490the thing is, i'm not. almost all models are trivially different variations of about a dozen different core transformer / state-space topologies. not sure why you're doing this shit.
giga brain move: have ai create a program for your task then have a grok bot or muse run that program/task for you on demand
>>110238483It's cheaper, especially at scale, to have a program to do something repetive.
I just ask sol to download me some porn whenever I want to jerk off
i just get gemini argon to roleplay as a gacha girl when i want to jerk off.
>saargon
yes sir. argon plays a mean hu tao/furina. nutted so hard
>>110238375cuckcoding general
>>110238500Yes there are similarities between models, what's your point? Each one still needs an implementation of the model graph, and the Rust crates provide a Python-free way of using the resulting compiled model. The refactor made all of them follow the same code style, have similar APIs, etc. The overall goal of the project is to have optimized models with no Python/PyTorch dependencies, for example, I can currently deploy Qwen-Image 2.1 or literally any model in a few MB. Developers can use the Rust crates to develop applications that use ML models without caring about the implementation of the model itself, they just use it, I am already dogfooding for my other projects, I use the qwen3-tts crate for something that loads a json manifest, generates the audio and writes out another manifest, all without Python/PyTorch slop and it's a lot faster, or another example uses ArcFace and DINOv3 to generate embeddings of faces and writes them to a database, again a lot faster than PyTorch, especially considering I'm on Windows with an AMD GPU.tl;dr you're a moron
>>110238543>being this localtardedhere's your one-man slop factory award!
>>110238536I am engineer saar the finest engineer in the west!
>cuckcoding general
>>110238543Use case for 800 AI models?
https://poal.me/61liikhttps://poal.me/61liikhttps://poal.me/61liik
>>110238543Nice hypothesis but I think futanari are not gay because they have a vagina still. Thanks for coming to my TED talk on the intricacies of the futanari species and you under vibe coding.
>>110238561There are many use cases
At least 800 of them, in fact.
as an enterprise paying 140% api rates it really pains me how expensive opus is
>>110238588
>>110238375https://www.youtube.com/watch?v=02MspGEWyuohttps://www.youtube.com/watch?v=02MspGEWyuohttps://www.youtube.com/watch?v=02MspGEWyuo
