Laut einer Studie steigern KI-Codierungsassistenten weder die Produktivität noch verhindern sie Burnout

It helps me type faster but thats about it. Most of my time Im blocked on people reviewing my stuff. It might make typing less tedious, but it doesnt matter all that much.

AI assistants are shrouded in smoke, making their value unclear.

For simple, repetitive, and common tasks, the tool should automatically handle these parts of the code, but for trickier, more complex requirements, most AI tools fail, making them unsuitable for the most stressful parts of software development.

It absolutely does help quite a bit with throwing code together, but it 100% requires someone skilled on the other end to refactor and clean up jank-ass code.

There is a more junior dev on my team that seems to use ChatGPT for *fucking everything*… he doesn’t realize I know, but recognizing GPT-output code is pretty simple if you’ve seen enough of it. Anyway… instead of taking the output and modifying bits and pieces of it to better fit within the needs of the application.. he just just copy/pastes it in, and if it works even a little bit: ship it. I reject practically 90% of his PRs, with the most recent one commenting essentially “see me after class” because it was so fucking bad.

I even went and asked GPT to solve that problem, and with some back and forth, got a reasonably workable solution that just needed some slight modification on my end.

IMO, AI is decent at getting a sort-of prototype up and running. It gives you a very low-fidelity version of code that might help you sort of get started…. but it is VERY MUCH not production ready… both because the implementation without a ton of back and forth is trash (or even better: personal touch-modifications that make it *truly* fit the application)… but – **and this is important** *-* ChatGPT-generated code is *fucking full* of security issues.

As soon as you’re out of education, the bottleneck on programming is not in turning ideas into lines of code.

It’s in understanding the technical and business requirements, and coming up with structures for getting things done that makes solving the immediate need easy, and the inevitable change less painful.

It’s in building a consistent model of the world that lets your code do what it needs to do, and makes communicating about needed changes easy.

It’s in constructing a platform that allows multiple devs with different skillsets and mental models to contribute according to their own strengths, and makes their task of doing the above easy for them.

If you’re a programmer, and you don’t think these are major limitations, I suspect that’s because someone is managing them for you, and you haven’t hit a hard edge yet. May you be able to overcome it when you do. I trust you have built ways of working so you are not stuck the instant your prompt engineering fails.

Edit: To wit, Burnout happens when management does not understand this, and treat programmers like factory workers. Productivity doesn’t increase because AI tooling does not address the bottleneck.

ChatGPT is great for writing boilerplate, unit tests, util functions.

None of those are the reason for burnout.

these must be really niche study cases; because ai coding definitely boosts productivity because i spend less time on good for my issues

I find ChatGPT to be good only for writing and explaining regexes.