2023-04-07 Me, meeting ChatGPT

#AI #ChatGPT #GitHub Copilot

My starting story with ChatGPT

Even though I heard it mentioned quite often, it took me a while to get around to try out ChatGPT myself. That was mostly because last year I played around with Stable Diffusion, whose language capabilities didn't seem particularily impressive to me, as phantastic as many of the generated pictures are. Seen in retrospect, I found my emotional reaction to ChatGPT pretty funny and interesting, when I finally tried it later. That's the story of that. And what I worry about and what I expect.

As a precursor, I started playing around with GitHub Copilot at the end of last year. Since I'm currently working on our open source projects Composum, that didn't seem troublesome from a data protection viewpoint. For a while that was just running in the background and made some suggestions now and then, but I didn't see much value in that. But it turned out that was because I used it wrongly. One day I had to write a semi-difficult method doing some tree traversal and processing things there. So I thought it'd be a nice thing to write the Javadoc documentation before I write the actual method, because I thought it'd help clearing things up in my mind when I do that right at the start. So I did, set the cursor into the method to start typing the code, and then went "What??" as suddenly a pretty sensible 8 line suggestion by GitHub Copilot appeared. I ended up coding my own thing, since it had misunderstood something, but that was pretty darn impressive.

A while later I was creating a JSP with a quick description of the Composum package manager, but didn't remember the exact HTML classes needed. So I thought: let's try that approach again and wrote a comment "Bootstrap panel with description". And lo, not only a correctly done bootstrap panel appeared - it was already prefilled with three paragraphs of nicely done text: The Composum Package manager is ..., You can download it at ... and so forth. Good enough to pass as written by a human. I went: "WHAT?"

After telling that a friend, he said that his job was secure, because he was working with the really old programming language PL/1, which is still a hard to replace part of some very old business logic. I thought: OK, let's put that to a test, and told ChatGPT to write a program in PL/1. And it did, and correctly fulfilled some requests to refactor the program, showing a pretty decent understanding of PL/1 (on which there is not much information or source code on the web) and also a pretty decent understanding of programming - up to a point, that is. I went: "WHAT?????"

After that, I was basically sneaking around ChatGPT for a month, again and again trying something and going "No, it cannot possibly be able to do that!". One problem was that in my mind neural networks still were basically a mathematical way to capture intuition - which they certainly were the way I used them sometimes for fun. I thought actual thinking with multiple steps was something rather in the far future. But as it turned out I was dead wrong, and it was already possible to train something quite close to human thinking into them.

As a kid, I basically wolfed down those stories about humanlike and superhuman computers / robots by Stanislaw Lem, Isaac Asimov etc., but I always thought that's maybe something my grandchildren would see, if I had any. But with that fundamental barrier broken, there is a pretty decent likelihood that I'll see quite a lot of that before even my retirement. With that breakneck speed their capabilities extended in the last few years, and now with people noticing and throwing big money after them, I'd rather expect seriously superhuman intelligence within 5 years.

That creates extremely mixed feelings for me. After all, we are living in a world where quite a lot of people will try to horribly abuse this to build the Terminator, and much of the rest will try to use it to replace all of their employees with machines, if possible, never minding that our society is still at a stage where paid work is the basis for living.

But, after all, this is also a childhood dream coming true. It's up to us what mixture of paradise and hell we create from that. I realized it's best to fetch the surfboard and try to ride that tsunami, and have as much fun with it as possible, and perhaps shape the future a little myself. At least it seems OpenAI managed to grow ChatGPT into something like a good-natured friend by character.

So, good luck to all of us!