But you still have to understand what you're doing, they don't produce accurate code all the time, and I'm not convinced they'll be able to just insert working code into existing complex systems.
I fully agree. I have been using Copilot for a while and the errors it makes in more complex domains (algorithms and data structures more complex than 'write me a quicksort') are just stupendous. Even at a more basic level, it regularly makes Rust syntax errors that my 10yo daughter has no issues spotting.
It has been really awesome for writing boilerplate code though. Especially with in-context learning, writing one example and then letting the model extrapolate it to other methods, functions, etc. works great and saves me a lot of time.
It's hard to predict the future. But the current best models feel more like a much better IntelliSense (though API hallucination is a serious issue) than something that is going to replace good programmers anytime soon (unless your task is writing boilerplate code).
I fully agree. I have been using Copilot for a while and the errors it makes in more complex domains (algorithms and data structures more complex than 'write me a quicksort') are just stupendous. Even at a more basic level, it regularly makes Rust syntax errors that my 10yo daughter has no issues spotting.
It has been really awesome for writing boilerplate code though. Especially with in-context learning, writing one example and then letting the model extrapolate it to other methods, functions, etc. works great and saves me a lot of time.
It's hard to predict the future. But the current best models feel more like a much better IntelliSense (though API hallucination is a serious issue) than something that is going to replace good programmers anytime soon (unless your task is writing boilerplate code).