weird take because it’s entirely dependent on context. what are you asking the ai to do, with what context, and what problem space.
if you’re asking an LLM to build you an efficient, secure, maintainable application in one shot I agree - it’s gonna produce garbage.
if you ask it narrow questions with intention context in not-too-obscure areas, like the same CRUD apps most people have been rebuilding for decades, it does quite well.
like most things, this isn’t binary.
Its more of a static analysis tool than a programmer replacement
replacement is a strong word. here are some things i’ve personally done with llm generated code. using the llm significantly sped up this process of this stuff, arguably by an order of magnitude.
- refactoring a bunch of shitty go code in a mature cli to stop using global variables and other codesmells. 100% generated in a one-shot, command by command
- fixed some small bugs in mature grpc api codebases
- built a personal finance dashboard
- implemented a PoC WebRTC for file sharing
I would just like to chime in with an aneqdote. I've used LLMs to learn how to code since gpt3 or so. I knew nothing about coding 2 years ago, took Harvard's CS50, then started doing portion scripts and things with the assistance of LLMs for my actually job which is engineering and truss design. Now it still takes a person with an eye for design and engineering to make something "good" but shortly after I was making python scripts I wanted to learn more. I not only used LLMs to assist in making stuff but the whole time I was asking about coding principles, reading, learning. I had it compare popular languages. Looked at everything from a first principals standpoint. I took a liking to Go. I've been using LLMs to help me the whole time. I work 45 hours a week still engineering my normal job. I have 2 kids and other hobbies. I went from knowing zero code to contributing to lots of open source freedom tech because I had the tools and drive to want to do it. The LLMs were the tools that have allowed me to learn and build on my own as fast as I have. I'm proud to say that in just a little over a week, I will have been working on grain, a nostr relay, for a full year. By myself mostly. Grain is no tech demo. I may not be professionally trained. I know grain is not perfect. I know it has flaws and I'm making it better all the time. But I'm arrogant enough to say I know it's pretty damn good for coming from one that only commits in his very limited free time that didn't know the first thing about go lang just over a year ago and only learned basic programming concepts 2 years ago.
Cc. nostr:nprofile1qqs9pk20ctv9srrg9vr354p03v0rrgsqkpggh2u45va77zz4mu5p6ccpzemhxue69uhk2er9dchxummnw3ezumrpdejz7qgkwaehxw309a5xjum59ehx7um5wghxcctwvshszrnhwden5te0dehhxtnvdakz7qrxnfk nostr:nprofile1qqsyprmrd0fxlnzl9xyfqva5gl9jgy0kp2cm3f0u3jecgtdtwk8aadgpz4mhxue69uhhyetvv9ujuerpd46hxtnfduhstmd26f because in this time I also contributed gonuts lightning backed for CLN too.
