are you guys using ai to code?i swear i tried, because everyone seems to be singing praises like crazy out of it, but it just seems to vomit tasteless code and i gave up trying to make it workidfk how everyone says it increases their productivity so much unless everyone is a codelet and im the only smart person (which is very very unlikely)
>>107556278Unless you're getting it to generate bog-standard boilerplate, you spent more time trying to tard wrangle AI to spit out something correct than you would have spend just figuring it out yourself.
>>107556278anotha cope thread
>>107556285boilerplate i most automate by writing macros or using template engines to spit out a basic version in most projects i join so i really dont get the benefiteven automated tests i feel its good to write yourself
>>107556290> idfk how everyone says it increases their productivity so much unless everyone is a codelet and im the only smart person (which is very very unlikely)
>>107556303i mean all the replies are going to be cope like this >>107556285unironically go to reddit for better discussion
>>107556325i feel like the autistic cracked programmers live here though
>>107556278>using AI to code>when you could be cooming
>>107556278Ye, mostly things I don't know or forget (how to use adapter X or API Y) but never ground work.
>>107556374Ah yeah, for searching or giving toy examples it works relatively well. For more than that (like autocomplete, code generation), it sucks imho.
>>107556362Teach me how to coom with ai, i don't get it.
>>107556596What kinda' cooms are you into? Wanna make your imagination come to life with genning AI smut?Wanna make your imagination come to life with roleplaying AI smut? Wanna act on your must perverse and borderline retarded fetishes so you burn out every last neuron and nobody will ever know? AI.
>>107556392Except VisualStudio C#'s offline ai autocompletion is pretty litBut anything more and it fucks up kekKnowing this, can you imagine how some vibe coded codebases must look like at Microsoft ?
>>107556708intellisense was always goated and one of the few things microsoft did wellbut yeah i can imagine how shit it is the vibe coded crap
>>107556278I tried. It is slop and takes more time troubleshooting. The experience was is even worse than trying to correct spaghetti-legacy code.
>>107556278yes its a great tool. you need to know how to code imo. its an assistant, are you too lazy to write nested loops for a function? ai can chunk it out no sweat. do you not know how to program? you're going to get dogshit results
>>107556325Nothing about what I said is cope. AI gives you the perception of increased productivity because you're outsourcing the hard part of thinking a problem through (and you're rewarded with an instantaneous result), but in real terms, you end up spending more time tard wrangling it to produce the desired result. For any kind of novel code generation, AI is a tedious pain in the ass. It can be useful to point you in the right direction, give you a rough draft for what you're trying to accomplish (sometimes), or explain a concept in more detail, but getting it to generate functional code (beyond, as I mentioned, boilerplate) that follows best practices and actually makes sense in the context of your project is a giant waste of time. I find that AI can increase productivity when you use it as a drive-by, first-line tool to get an immediate, direct response--but when you spend time dwelling on its retarded hallucinations and half-functional code, you're just hurting yourself in the long run.
>>107556278i use it for boilerplate and working examples for shit im learning atmotherwise its like you said-its a chore to get the code it produces to workmaybe you can get it to work better by basically having a sort of "header" with all the relevant info in the system prompt could improve things.its an idea i had, didnt try it out yet
>>107556865nested loops are kind of a code smell unless you’re doing a problem that is really o(n2) (which are rare)
>>107556278It works fairly well for me. It's best suited for routine CRUD/business logic/forms but honestly for most projects that's probably 90% of the code you need to write. You'll eventually develop an instinct for what you can use an LLM for vs what would be better to write yourself.Break down your prompts into small features and break those down into steps. After each step, review the changes and commit them (can always squash later).Point it to similar existing code that it should refer to and match, like>see foo.ts and bar/*.ts for examples, then write your output in baz.tsUse the best models (opus 4.5) with good high-token-usage tools, like claude code or cursor max mode. It'll get expensive obviously.
>>107556870i think its good for researching and when you don’t want to spend like 2 hours to relearn how to implement a red-black tree or some super specific problembut even for boilerplatey stuff it kinda sucks, idk, the unit tests it spit out for me kinda sucked
>>107556278I hate coding in Python. I view it as a jeet slop language. There's no honor in writing Python shitcode, but sometimes it's necessary because I want to run some TTS model or something that only has a Pytorch implementation. So I just ask AI to shit out a gradio web interface or something like that.That's pretty much the only use case for AI coding, at least for me.Actually Deepseek is pretty good at Powershell as well. By looking at what Deepseek shat out I was able to learn powershell to a basic degree. But I still never copypaste what it wrote verbatim due to the potential for data loss.
>>107556918i rarely write crud, im in mobile(ios)crud in mobile is maybe using sqlite and you don’t get ORMs or similar shit for that OR being a json pretty printer if you are doing something super simpleanything that is a bit more complex (searching is a classic example) needs full attention because you’re writing concurrent code
>>107556923unit tests?nonono anonthe boilerplate'est boilerplate that ever boilerplatedunit tests are extremely important and are heavily dependent on the actual algoits DEFINITELY something you want to write yourselfyou can think of it this way:ai gen unit tests are like a criminal who would investigate himselfif the machine thinks its code is okits unit tests will probably find the sameanother way of thinking about this:the unit tests need to be the highest quality and reliability because its the measuring stick you use to find errors in the rest of your code
>>107556957Yeah, which is why I said they suck. To be fair I’m also not a fan of TDD because I like to evolve my solutions from prototypes and I write tests after I think they look reasonable.
>>107556923>>107556957 contwhat could work pretty well i think, is once youre done, you still send your stuff to a chatbot, to get a "second set of eyes" so to speak.its not reliable, it will find false positives, but once it finds a positive positive, youre in the green
>>107556972Yeah, I am more excited for AI to do “semantic aware reviews” and feel it’s more useful for than than for writing code.
>>107556914thats another thing its really good for, taking my slow discrete simple simon method and turning it into a fast continuous O(1) calculation
>>107556278>>107556918 (Me)Just to add to this... generated code quality will normally be worse than what you would have written yourself. At least on its first pass. There'll be useless comments, more duplication, less structure. It might overlook utility functions that it should have used and write worse versions itself. Etc.I tend to be OK with that. It's a fair tradeoff for the savings in time and mental energy. You can tell it to make specific improvements, or refactor it yourself, or just accept it as-is if you're in a hurry.>>107556947sqlite should be no problem. The web project I'm working on doesn't use an ORM either, just postgres directly. I like to define the tables myself and get the LLM to generate corresponding typescript definitions and CRUD queries.
>>107557001tbf you sound like you dont know how to code
>>107556964classical tdd is kinda retarded to be honestits a specific method and order of doing thingstesting often is very important though>prototypes, testing once a solution is foundi work the same waybut i test prototypes too, this way i check for conformity with the expectationsits just that i dont test the limits of my algo, i check for functionality, thats how i know the prototype is conform to the expectationsonce i get a solution down, i do a full suite of testsi check the limits of my algo (0 size buffer, big buffer, etc, whenever a limiting clause would be tripped)i check my error catching mechanisms with deliberately wrong input (malformed data, erroneous data)but i test really often, i test pretty much everything.the only things i dont test are wrapper functionsbut then i write in c and its kinda part and parcel of using the language
>>107557010i often deal with a tiny wrapper to the C api though (either GRDB.swift or Sqlite.swift). Swift isn’t a dynamic language (it’s very much like Rust that prefers generics everywhere) so you can’t really use runtime trickery like JVM languages or Go do, so you don’t get those reflection tricks,
It's trained on all the code they could find and most code out there is trash. It's mostly fine for immediately obvious boiler plate code.
>>107556278I use some ollama coder and all it provides is examples on how to use some libraries in C, it does save time a lot but it will never complete any project.
>>107557028I don't use reflection tricks either. Point it to your schema and to some existing examples and you should have no problem getting it to generate types/queries/wrapper functions.
