Piero Bosio Social Web Site Personale

@catsalad
Of course!

One coworker -- who I'll call Xavier -- does everything through LLMs. He's the kind of developer that managers who have never been programmers adore: 3000 lines of code per day.

I just realized that Xavier cannot read code, even code that he submits for review. He "understands" code by running it against test files and seeing whether results are reasonable. He can only say what the code does, not what causes it to behave in a way.

But that has several problems. An obvious first issue is that if a problem doesn't show up in the test file, then it will never be fixed. A less obvious issue is that his code is brittle and it generalizes very poorly.

Because Xavier doesn't read code, he has a very tough time imagining "What might go wrong?" And because he relies on the LLM, he misses very broad solutions, like using well-established libraries that solve dozens of problems at once.

Programmers who dive deep are still very, very useful.