The “catastrophe scenario” of AI in the software engineering job market has gained quite a lot of popularity, with people pointing out that current models already “do a better job” (what does that even mean?) than many professional programmers. I’m not sure I completely buy into this.

I’ve been playing with dalle-mini the last few weeks. Part of what makes it fun to play with are the bizarre and obtuse outputs. They reached that sweet spot between laughably bad and frighteningly perfect: they’re good enough to be understood and enjoyed, basically.

I think that incompleteness is part of what makes it so amusing to toy with these things, and conversely what will make future versions much less fun.