> for lots of tasks, the most precise language is the programming language
This is my problem with AI, or "vibe coding" or whatever you want to call it.
We already have many language(s) for expressing what we want the computer to do, and I find that natural language is the most difficult one to accomplish that goal.
Sure, LLMs might be able to whip up a fancy landing page and some basic API really quick, but when it comes to solving harder problems I find it incredibly frustrating and difficult to try and prompt in English, when the programming language I'm working in is already well suited to that task and is much more explicit in what I want.
Maybe I'm using it wrong, but it's way less cognitive overhead for me to just type for for x,y := range something {} than it is to try and prompt "Iterate over this list of values...etc."
I've found that the only prompts that actually work for generating code reliably are the ones where you already know exactly what code it will output -- where nearly every part of the "how" is precisely clear, it just needs to be compiled from human language into code. Writing this perfect prompt often turns out to be a waste of time. It's a classic specification problem. Human languages will let you under-specify a problem, but the result will not be pleasant.
Maybe I'm using it wrong, but it's way less cognitive overhead for me to just type for for x,y := range something {} than it is to try and prompt "Iterate over this list of values...etc."
I'd say that prompting "Iterate over this list of values...etc." is definitely using it very wrong (autocomplete should more or less handle that sort of thing anyway). Prompts should be more in line with "write a C++ function that can parse XML files that look like this (upload a few sample files) and return the data in this struct (copy and paste struct from your header file)" followed by "write a set of unit tests for this function". You then check that the unit tests look reasonable, add any other things you feel you should test for, make sure the generated code passes the unit tests, and then check it in.
This is my problem with AI, or "vibe coding" or whatever you want to call it.
We already have many language(s) for expressing what we want the computer to do, and I find that natural language is the most difficult one to accomplish that goal.
Sure, LLMs might be able to whip up a fancy landing page and some basic API really quick, but when it comes to solving harder problems I find it incredibly frustrating and difficult to try and prompt in English, when the programming language I'm working in is already well suited to that task and is much more explicit in what I want.
Maybe I'm using it wrong, but it's way less cognitive overhead for me to just type for for x,y := range something {} than it is to try and prompt "Iterate over this list of values...etc."
When I'm programming, I'm not thinking in English