WARNING ! AI rots your Brain

My personal experience prototyping with AI aid often feels like my brain is exploding. This is because, while beneficial in some respects, the process proves quite exhaustive, demanding constant redefinition of specifications, multiple corrections, and frequent rephrasing of prompts to prevent hallucinations, sometimes even requiring a complete reset of the interaction history. I've observed this to be mentally more taxing than developing from scratch, carrying a significant risk if the generated code isn't meticulously verified. Future AI iterations ideally need better context retention and more robust error detection to reduce such extensive manual oversight, especially as complex logic still heavily relies on human expertise.

My personal experience prototyping with AI aid often feels like my brain is exploding. This is because, while beneficial in some respects, the process proves quite exhaustive, demanding constant redefinition of specifications, multiple corrections, and frequent rephrasing of prompts to prevent hallucinations, sometimes even requiring a complete reset of the interaction history. I've observed this to be mentally more taxing than developing from scratch, carrying a significant risk if the generated code isn't meticulously verified. Future AI iterations ideally need better context retention and more robust error detection to reduce such extensive manual oversight, especially as complex logic still heavily relies on human expertise.

At this moment, I think it is more useful to find the relevant APIs, and get example code of how they are used.
Trying to let it build a working (demo) application, by chatting: no.

(Typing from mobile so have to be brief) Which tools do you use? I'm using paid CoPilot although I think there are better tools now. 

I don't know how much it retains, but I find it useful even at drafting (based on my comments) the function prototypes or class header files (and guiding it yourself if it's not what is expected, although sometimes it's interesting to see where it goes for a little bit, in case I've missed the woods for the trees), and then it seems to remember the methods and member variables when it comes to the source code, so that it's probably focusing effort on small functions rather than ingesting large volumes of code for generating its output (I'm only guessing though since it's a black box).

More often than not it also learns to output things in the current code style (e.g. my variable naming conventions) but if it sometimes picks something else, I can nip it in the bud right at the header files, then it remembers that for the rest of the code.

I've been using Copilot at work, but honestly, it's been pretty frustrating. I'm a big fan of Gemini instead.

For SQL databases, Gemini is absolutely brilliant – it just takes all the boring work right out of it. With Java, it's really good too, but you do need to be clear about the level of abstraction and the design patterns you want so you don't just get generic, copy-pasted code.

I've noticed huge improvements in its VHDL and SystemVerilog capabilities in recent updates. It's fantastic for generating testbenches, which is super helpful, though I do have to put in extra effort to verify its output. I'm also studying RISC-V processor design in SystemVerilog, and Gemini has been a lifesaver for quickly generating small test programs for the RISC-V instructions I'm implementing

That's very useful feedback, I'll give Gemini a try at some point.