Unlike most computer books, Hans Moravec's Mind Children: the Future of Robot and Human Intelligence has legs; it's still readable and relevant today, 15 years after it first appeared. Moravec's book is so good because, at its best, it focuses not on narrow computational specifics — all ephemeral and essentially uninteresting — but on key ideas associated with computer science and artificial intelligence (AI).
Moravec asserts several critical points:
- philosophy — what minds can do, computational machinery can do
- hardware — exponential growth in processor speed and memory size is likely to continue long enough for human-class raw computational power to be cheap and commonplace within a few decades
- software — research in robotics, to solve the challenges met by real-world mobile machines, is the best way to make progress in developing AI
These notions are all debatable, of course ... but Moravec puts forth strong arguments for his side. On the robot research front he's admittedly self-interested, as head of a robotics lab at Carnegie Mellon University. His theories go back many years and show the fingerprints of collaborators including Vernor Vinge and Robert Forward, as he is happy to acknowledge. The book throughout is infused with contagious energy and enthusiasm.
A personal footnote: back around 1976, when some essays that form parts of Mind Children were taking shape, I met Hans — in a peripheral and virtual way. He was a seasoned graduate student at Stanford; I was a relatively new grad student at Caltech. Somehow (details forgotten) we began to correspond via email over the Arpanet. This was rather unusual at the time, given the small universe of people on the network and the barriers to finding one another. On a DECwriter terminal in the high energy physics building I dialed in on a few-hundred-baud line, typed commands to a remote computer, and read the results on dot-matrix paper printout. Most of what I did was ostensibly related to my dissertation work, but in between checking my equations on MACSYMA I also managed to find time to study FORTH compiler/interpreter architecture and, sporadically, exchange a few letters with a few other ur-netizens. Moravec was one such. He was tolerant of my naïve questions and kind enough to send me a copy of his early article on how a living, thinking brain could, neuron by neuron, be transferred into a computer. Heady bedtime reading; thanks, Hans!
Moravec does a fine job in Mind Children of sketching out advanced algorithms and then exploring their philosophical implications. One striking example from his final chapter: "Hashlife" (or "Hash Life"), a brilliant approach developed by Bill Gosper for running Conway's cellular automaton ("Life") at hyper speed. It's a bit too hairy to describe here; as Richard P. Feynman said to some newspaper reporters about his Nobel-winning work, if it were simple enough to explain in a few sentences then it wouldn't be worth a prize. Moravec covers Hashlife at a general level in four pages with a couple of apropos diagrams, just enough to get the key ideas across. Then, as compactly, he applies the Hashlife metaphor to our universe. Mind-expanding ...
But how does Moravec's central thesis — that we'll see human-like machine intelligence by 2030 — stand up today, over a third of the way down his timeline? Perhaps it's too early to say, but my hunch is that Hans's hubris is showing. Even an exponential curve, which starts off slow as it accelerates at an ever-accelerating pace, should have made at least a bit of significant improvement by now. We've gotten bigger, faster hardware aplenty, along with high-resolution video games galore. But where's the fundamental software progress? I remain a short-term skeptic, even if I'm a long-term believer in strong AI.
(see also GeniusAndComplexity (25 May 1999), IntelligenceAugmentation (25 Aug 2001), VernorVinge (17 Sep 2001), FastForward (21 Feb 2002), FasterForward (26 Sep 2002), DeadBeginnings (28 Sep 2002), ...)