Funny, I remember the first time I had this conversation. I had just graduated, around 1983, and we were worried about CASE tools taking our jobs.

And every five or ten years there was a new flurry. And I am still employed.

We'll get there one day. But Copilot is the wrong direction.

We don't have strong AI or anything like it. The programs that generate the code are machine learning. This is an amazing technology, but it's only about extracting statistical patterns from a huge corpus of existing data.

To write programs correctly, you can not simply do what the average programmer did in a somewhat similar situation!

You need to actually have a model of the state of the program in your head.

I've taught a huge number of students. This idea, that you can write programs "by feel", just never works. At some point, you have to understand the program.

Is that possible for an AI program? Sure, why not?

But Copilot is nothing like that, and there's not even a theoretical way to take an ML model and turn it into an AI model that understands concepts like variables, scope, iteration, and that sort of thing.

And don't forget the fact that 90% of the code that Copilot is based on is bad, because of Sturgeon's Law, and Copilot has no way to tell the good from the bad.

(The one person I know who used Copilot was writing on the second day that he was using Copilot to fix the bug that it had put in on the first day! He thought that was cool. But I get paid good money because my programs actually work the first time.)