Hacker News Reader

Comment by al_borland | Hacker News Reader

I spent the last half of my day today trying to have AI write some boilerplate for me, to basically port something I already wrote to a different platform. I tried 3 different models and it fell on its face repeatedly. I probably would have made more progress if I didn’t use it at all. That will probably be the path I go tomorrow.

I think you’re falling for some of the hype. I’m not feeling the magic others often talk about.

View on HN

AI's capabilities are very unevenly distributed. If there is a wealth of quality information on the internet about something, then it usually comes up with a solid answer. But where information is scarce it often just bullshits its way. Two examples:

Trying to get some JS code to work for a GUI - mostly correct, but I did cross-check with MDN.

Tried to debug PXE booting Raspberry Pi from Haiku-OS as a server. A mixture of good information on PXE as per the standards, but sketchy when it comes to Raspberry Pi's non-standard implementation. Failed to solve problems arising from Haiku-OS beta5 state.

As career advice: skate to where the puck will be! Get proficient at solving domain specific and edge case scenarios. Fixing AI created monsters is going to be like tattoo removal clinics erasing regrets of past bad choices.

This is the big issue with AI in the real world. It does great in tech demos, I got it to one shot a password generator site with some customizations I wanted that most don’t have. However, when it comes to my actual job it’s almost always weird stuff done in a weird way that there isn’t much information out there for. I have to assume a lot of companies are this way. Enterprise apps tend not to have a bunch of code examples out on GitHub for AI to train on.

A couple years ago we got some professional services from a well known vendor that makes the platform we were running. What they developed was basically something you’d see in a blog post, it was useless. If that’s what AI is going to do, it’s not helpful. We ended up throwing away all their code and doing our own thing so that it would scale and not be a pain to use.

As far as I can tell this will be a weakness of AI for some time to come.

Reply on news.ycombinator.com