It's strange: we built computers that beat chess champions, crack calculus, prove theorems, and diagnose disease better than doctors — yet we still can't reliably get one to recognize a coffee mug on a cluttered desk or walk across a room without falling over. For a long time that looked completely backwards.
The early AI researchers fell for it. The "hard" problems turned out to be the easy ones, because humans invented them. Chess, calculus, theorem-proving — these come with small, well-defined rules and a tidy problem domain. A computer with very little horsepower can apply a known set of rules to a narrow space and look brilliant. The human experts handed it the intelligence.
The genuinely hard problems are the ones that feel effortless to us: identifying an object by sight, picking it up without crushing it, staying upright while walking. We never "invented" seeing or walking, so there's no neat rulebook — just staggering amounts of computation. Basic edge detection in the human eye alone takes something like trillions of operations a second, and the serial computers of the early AI era weren't remotely close.
So progress runs in reverse. We started at the abstract end — theorems and chess — and we've been working down toward the primitive ever since: reading text, recognizing speech, and only now, with enough hardware and a half-century of statistical tricks, starting on vision, movement, and plain common sense.
What gets me is that evolution solved the hard problems first, billions of years before we showed up — vision and locomotion are ancient, language and math are brand new. That's the difference between designing bottom-up and top-down. And it's humbling: the part of your brain that recognizes a face or catches a falling glass runs whether you like it or not. The conscious, "smart" part you're so proud of has almost no say in it.