Comment by gregfjohnson

This all reminds me of one of the most foundational and profound papers ever written about software development: Peter Naur's "Programming as Theory Building". I have seen colleagues get excited about using Claude to write their software for them, and then end up spending at least as much time as if they had written it themselves trying to develop a theory the code that was produced, and an understanding sufficient to correct the problems and bugs in the created code. Every professional software engineer confronts the situation of digging into and dealing with a big wad of legacy code. However, most of us prefer those occasions when we can write some code fresh, and develop a theory and deep understanding from the get-go. Reverse-engineering out a sufficient theory of legacy code to be able to responsibly modify it is hard and at times unsatisfying. I don't relish the prospect of having that be the sum total of all my effort as a software engineer, when the "legacy code" I need to struggle to understand is code generated by an AI tool.