I view these LLM tools as just another code generator. I have no idea how many bespoke single-use code generators I've written, but I know that I've spent a non

I view these LLM tools as just another code generator. I have no idea how many bespoke single-use code generators I've written, but I know that I've spent a nontrivial amount of time writing code that writes code. If an LLM can do it with a natural language interface, it'll save me a lot of time. But also, I would never trust a script I threw together in 15 minutes to actually produce real code and solve the problem. All it does is generate the text I tell it to. It can't understand the system or how it works, it just procedurally spits out text. In the same way that a dozen lines of Python cannot understand your program, an LLM is also fundamentally incapable of understanding it. That's the crucial part of programming. An LLM will give you text all day, but it can't write your program. Sure an LLM can produce trivial scripts, but in my experience so far, it can only reliably generate trivial programs that I could write blindfolded in the same amount of time. If we just treat these tools like the text generators they are instead of insisting they're code generators, we'll all be better off.