In the past few weeks, I got to observe a few people on the verge of addiction to AI: spending tokens on agents and anxiously wait for the answer as if they were slot machines — and never knowing when to stop.
I must say that coding with agents is a lot of fun (someone was comparing to playing video games), since you get to write down intentions and get some working results without being stuck by bugs or the dread of having to scroll through a dozen files and lose your focus. And this true too for my students, who have been able to make rapid progress in a little time: dodging roadblocks prevents them from sapping their enthusiasm, and they can go quite far with proper guidance. The downside of course is that you end up with code that is very difficult to maintain, a problem identified by Donald Knuth: “premature optimization is the root of all evil.”
That is particularly true in science, where you don’t really know where you’re going, and you should be able to rapidly iterate on ideas. I have a special hatred for python, a language that is not made for science (numpy and matplotlib are necessarily inferior to Matlab, jupyter notebook are cute but a terrible mess and error messages are difficult to parse) – yet it is the only game in town (thanks to open source packages) and I’m sorry I have to impose it on the people who work with me.