I find the juxtaposition of AI being so good that we no longer need to look at the code, and it being so bad that we can't expect it to write maintainable code, so very funny.

Like, I would expect an advance AI that can load, read, and comprehend (whatever it means) entire systems, along with all related documentation, to also be able to create perfect abstractions that accurately model the problem space. And do so continuously.

Why is it not an expectation that an AI, in addition to superhuman coding speed, should also have superhuman software design capabilities? What's even the point of it being able to hold entire systems in context if it's not gonna use this power to infer, create, and maintian the perfect design?

When I think of a system designed by AI, I imagine mesmerising fractals, the music of spheres, a perfect engine. Not a blob of slop that AI can maul into desired shape because it's got the biggest club.

@alexozun I wholly agree with you but I don’t know if that is due to my own ego and a desire to believe I still have value.

@alexozun If we are to get to the point of generated code being sustainable, usable, and trustworthy, then that code will necessarily have to be well-structured.

It won’t necessarily have to generate code comprehensible to humans, but that code will need to map well to the requirements set out by humans.

@alexozun Just as a compiler turns (valid) human-written code into well-structured machine code that most humans would not understand, these new code generators may eventually be made functional enough to generate well-structured code that most (or even all?) humans would not understand (whether in high-level languages then processed by compilers, or even in machine code directly).
@wtfSwiftUI Absolutely! That's exactly how I expect AI-generated codebases to be designed and structured. Optimized not for human legibility but for correctness in a broad sense. Which doesn't just mean "works as expected", but also "built as expected", in accordance with all sorts of meta-requirements around the SDLC.
@wtfSwiftUI For example, if AI can infer from the history of changes that I like to update button colors every week, it must pro-actively put accurate abstractions in place to optimize future changes. And it must revisit and evolve this design regularly.