AI is whatever machines can't do yet.

The goalpost keeps moving

There's something almost funny about this definition—it suggests that AI is just a moving target, constantly shrinking as technology advances. A task feels magical and "intelligent" until the moment we build something that does it reliably. Then we collectively shrug and say, "Well, that's just code." Chess computers stopped being impressive the moment they beat Kasparov. Now we barely notice when our phones correct our spelling or suggest what we meant to type.

This matters today because it forces us to think differently about what we're actually anxious about with modern AI. We tend to imagine intelligence as some unified thing—like a smarter version of us that will eventually arrive fully formed. But Tesler's point suggests intelligence isn't a singular achievement; it's a series of specific tricks we haven't automated yet. The things we call intelligent right now (like language models) might just be the ones we haven't quite figured out how to commodify and simplify. In five years, we probably won't call them intelligent anymore.

The real insight is humbling: what we need to worry about isn't AI becoming conscious or superintelligent in some abstract way. It's that useful, narrow tasks keep getting automated faster than we adapt. The goalpost for "real" intelligence keeps moving because intelligence itself is less mystical and more just "the stuff we haven't optimized yet."