I've found it can just about be useful for "Here's my data - make a schema of it" or "Here's my function - make an argparse interface". Stuff I could do myself but find very tedious. Then I check it, fix its various dumb assumptions, and go from there.

Mostly though it's like working with an over-presumptuous junior. "Oh no, don't do that, it's a bad idea because security! What if (scenario that doesn't apply)" (when doing something in a sandbox because the secured production bits aren't yet online and I need to get some work done while IT fanny about fixing things for people that aren't me).

Something I've found it useful for is as a natural language interface for queries that I don't have the terminology for. As in "I've heard of this thing - give me an overview of what the library does?" or "I have this problem - what are popular solutions to it?". Things where I only know one way to do it and it feels like there's probably lots of other ways to accomplish it. I might well reject those, but it's good to know what else exists.

In an ideal world that information would be more readily available elsewhere but search engines are such a bin fire these days.

I was in the same boat about...3mos ago. But recent tooling is kind of making me rethink things. And to be honest I'm kind of surprised. I'm fairly anti-AI.

Is it perfect? Fuck no. But with the right prompts and gates, I'm genuinely surprised. Yes, I still have to tweak, but we're talking entire features being 80% stubbed in sub 1 minute. More if I want it to test and iterate.

My major concern is the people doing this and not reviewing the code and shipping it. Because it definitely needs massaging...ESPECIALLY for security reasons.

I mostly use AI as advanced autocomplete. But even just using it for documentation is wrong so often that I do't use it for anything more complex than tutorial level.

I got pretty far with cursor.com when doing basic stuff that i have to spend more time looking up documentation than writing code, but I wouldn't trust it with complex usec cases at this point.

I check back every 6 months or so, to keep track of the progress. Maybe I can spent my days as a software developer drinking cocktails by the pool yelling prompts into the machine soon, but so far I am not concerned I'll be replaced anytime soon.

I don’t honestly believe that AI can save me time as a developer. I’ve tried several AI agents and every single one cost me time.

I have had the exact same experience many times. But I just keep trying it out anyway, often with hilariously bad results.

I am beginning to realize that I like cool technology more than I like being productive.

AI can absolutely save you time, if you use it right. Don't expect it to magically be as good as a real programmer... but for instance I made an HTML visualisation of some stuff using Claude, and while it got it a bit wrong, fixing it took me maybe 20 minutes, while writing it from scratch would have taken me at least a couple of hours.