i think its pretty cool. i think its awesome to watch the system figure out how to fix problems (though, i'd be scared to death of how it tries to fix problems in bigger codebases). but this isn't anywhere near being affordable enough or trustworthy enough to use for anything serious - ThreadSky

hailey.at • 17 hours ago

i think its pretty cool. i think its awesome to watch the system figure out how to fix problems (though, i'd be scared to death of how it tries to fix problems in bigger codebases). but this isn't anywhere near being affordable enough or trustworthy enough to use for anything serious

Comments

tyranny.sparklenoise.com•17 hours ago

it’s a fancier advanced version of using ruby to scaffold a database in 2005

tyranny.sparklenoise.com•17 hours ago

“do this thing so I can more perfectly form it”

hailey.at•17 hours ago

also, ngl code completion from ai is boring as fuck. i wanna type on my mechanical keyboard, not press tab.

dane.computer•16 hours ago

both times i've tried it (copilot and the zed ai autocomplete) I've turned it off. it autocompleted what I wanted maybe 40% of the time which was getting annoying. also my new keyboard sounds good as fuck and i love the clacking sounds

dogpawhat.tech•17 hours ago

Yeah the waiting is boring as fuck rn

domi.zip•17 hours ago

you have to wait?

dogpawhat.tech•17 hours ago

Read the time section here https://per.blog/2025/02/16/my-llm-codegen-workflow-atm/

3.7 can take a while if you give it a complex prompt and leave it on agent mode

dogpawhat.tech•17 hours ago

It technically could write code a lot faster them me but I'm really not convinced leaving my brain to wander while it does this is the best use of all the compute inside and outside my body.

Also Cursor Agent really needs notifications jesus. I'd tag them here but there still on the Bad Place

mhlakhani.bsky.social•16 hours ago

I use it to get fairly repetitive work done on the side that’s mindless and I can audit while it runs; while I work on something else. That’s done wonders for me

domi.zip•17 hours ago

Oh, those are super complex queries. I don't really do those as I see no point of having LLM do the fun stuff of creating apis and architecture of the code. I have it write boilerplate or ingesters/types for input files/apis. Usually instant stuff even on smaller, local models.

29sunset.com•17 hours ago

Posted this about AI aiding serious research, but guessing it's applicable across many fields

ripperoni.com•14 hours ago

It’s also just not where most of development time goes in my experience. Historically if boilerplate is used so often, you just make macros or templates to take care of it

calayucu.com•15 hours ago

do you mainly use it in a chatbot setting then? i.e. typing out prompts / iterating in a "conversation"

that mode seems slow (it's probably faster to write the code than to prepare a good prompt) but maybe I'm missing other modes of interacting with these tools

hailey.at•14 hours ago

i mostly just use it to rubber duck ideas and do some "generic docs searching" (claude and gpt are both really good at finding doc info, since they're just trained on all the public info like docs sites as well as open source examples). otherwise, i dont use copilot or anything similar

ave.zone•14 hours ago

ime they still tend to hallucinate features not in the actual library when I try to find something obscure or nonexistent, so without network access, it's hard to rely on them for doc info

mary.my.id•14 hours ago

theyve been really useful for this, what you save time on is writing code for the prototype iterations, then you'd do the finishing touches since it's impossible to fit every criteria into a prompt

gregdibartolomeo.bsky.social•17 hours ago

💯 Agreed. Everything I know about how generative AI actually works, I would not recommend using it fully for code work.

Can it help you as a tool? Sure. Can it save a little time? Sure.

You will still need a professional educated human coder to verify that the output is correct and useful.

hailey.at•17 hours ago

the biggest problem is that there's definitely some "wow this actually looks pretty good!" bias that you can fall for (i've fallen for this) where because it _seems_ to be writing good code, you just assume it _will_ be good code

hailey.at•17 hours ago

particularly if it worked the first 10 times you used it. that 11th time might be a colossal fuckup, but it's too late because you've learned to trust it.

cbvt.bsky.social•15 hours ago

The sinister aspect of something that is almost right is that it fails in much more subtle ways

gregdibartolomeo.bsky.social•17 hours ago

Yup, it’s all fun and games till you get an output that fucks everything up and then you will never trust it the same way again.

It’s extremely difficult to gain back trust once something shows as “occasionally wrong”.

aendra.com•15 hours ago

There's a second order effect to this when you review your colleagues' code. When you learn to trust a developer by reviewing their code a lot, but then they make you suddenly review sloppy AI code, you will always be more careful reviewing their code in the future thinking it might be AI-generated.

ripperoni.com•14 hours ago

The thing about AI code that’s truly the worst, is that bad or novice developers will use it to generate code they don’t understand. That’s when it’s really at its worst

Comments

Posting Rules

Reply