When people say “taste,” what they actually mean is experience. Pattern recognition built up over years of doing the work. But calling it “taste” instead of “experience” does something subtle and harmful: it makes a learnable skill sound like a gift.
The tooling around AI should be to improve the quality of the programmer. Not to write the code for the programmer.
For example if you ask an agent how to scale things well, and best practices in architecture, it will have a lot of resources on it. But that does not mean the code it will produce when you ask it to write a programme will consider and include the best practices it gave you in a separate question. That is the ‘intelligence’ part that LLMs cannot have. If you ask a it to do a certain way it will create it. Context tries to address this by prompting the user to give more, but that is not persistent.
This is exactly why senior devs finding LLMs works for them, because they know ‘how’ to do it, and they explicitly state it. But at the same time junior devs feel they think the code written by LLM is the ‘best’ way so solve a problem and superior in quality, even if it is not, because they don’t know any better.
Tooling should be able to help the developers improve their knowledge and skill on ‘how’ to do it. Instead it always focus on writing the code. I want to add that I’m not talking about algorithms. But every aspect of coding, in which the programmer needs to know ‘how’ to do it.
Absolutely correct, and likely the eventual outcome (tool not replacement), because LLMs are not, and very likely never will even be close to AGI (let alone techbro dreams of ASI). Code smell is a thing, and LLMs reek. It is possible to learn from codegen, but that sure as rut is not it’s purpose. mostly an excuse to drive down wages to make the rich richer.