Books

Short works

Books : reviews

Hans Moravec.
Mind Children: the future of robot and human intelligence.
Harvard. 1988

rating : 1.5 : unmissable

Mind blowing!

Hans Moravec.
Robot: mere machine to transcendent mind.
OUP. 1999

rating : 3 : worth reading
review : 10 March 2005

Robot is a follow-up to the earlier Mind Children. That had already blown my mind a decade or so earlier, which is possibly one reason that I didn't find this incarnation quite so thrilling, although certainly worth the read.

Moravec charts the history of intelligent robots, from the early research days of AI to today, then on to tomorrow and the distant future. His thesis is that artificial intelligence is merely a matter of more computer power and cleverer algorithms for path planning and vision recognition, and backs this up with graphs of improvements of various devices, including autonomous vehicles. He provides a timescale for humanlike competence. But there's a problem with it.

pp.vii-viii. [in 1978] I argued ... that humanlike performance needed millionfold greater computer power, but might be attained in a crash program in as little as ten years by interconnecting millions of then-new microprocessors. ... [That] article implausibly called for someone to invest billions of dollars in computer hardware to possible produce one humanlike machine. Updates of the article were more sedate, offering human competence in multi-million dollar computers in twenty years. In 1988 ... humanlike performance was rescheduled for ten-thousand dollar personal computers in forty years. ... this book projects humanlike competence in thousand-dollar computers in forty years. A slight rise in the estimated difficulty has been partially offset by faster growth in computer power.

So, it's getting progressively cheaper -- but no closer! There's a standard joke about AI (and also fusion power) that it's twenty years in the future ... always has been, always will be. It seems like the AI joke needs updating: it's actually receding!

Moravec points out that some of the reason for this lack of progress is that for several decades there was continually-reducing funding for AI research. Improved computer power as a result of Moore's law merely balanced this drop in spending power, and so the community had essentially constant-power machines available to them.

p64. intelligent machine research did not make steady progress in its first fifty years---it marked time for thirty of them! Although general computer power grew a hundred thousand fold from 1960 to 1990, the computer power available to AI programs barely budged from 1 MIPS during those three decades.

However, this plateau is now in the past, and the power is increasing again, so progress should be swift. Moravec lays out a timetable for the next century. I unfairly quote the first prediction (note the date of publication, versus the date of review).

p92. Utility robots for the home could then follow, maybe around 2005, drawn by an assured mass market for affordable machines that could effectively keep a house clean.

Unfortunately, no mass-market home robot cleaners in sight as of March 2005. Part of this inaccuracy is no doubt due to the usual problem of prediction: exponential growth means that advances are a lot slower than expected in the short term, but a lot faster than expected in the long term. So the rest of the timetable may fare better. But I don't think so. Moravec appears to be claiming that all that's needed for intelligence is more power and better, but still "good old-fashioned AI" algorithms. He backs this up with a story about Kasparov's struggle with Big Blue:

p67. Kasparov ... claims to see into opponents' minds during play, intuiting and exploiting their plans, insights, and oversights. In ordinary chess computers, he reports a mechanical predictability, stemming from their undiscriminating but limited lookahead and absence of long-term strategy. In Deep Blue, to his consternation, he saw instead an "alien intelligence".

But this is just chess, a closed system with well-defined rules. In more open systems, like robots moving in cluttered environments, performance is improving, but Moravec still seems to want mere speed-up of the path planning process, rather than an all-together different approach. Why should a robot want to plan its path in detail before it moves? The old proverb springs to mind: "He who deliberates fully before taking a step will spend his entire life on one foot." And in an open, dynamic environment, full of other entities also moving, pre-planning is not an option. Moravec does discuss some other approaches, such as Brook's subsumption architectures, but gives them short shrift.

However, whether one believes in the precise time-table and algorithmic approach or not, it is reasonable to assume that humanlike robot competence will be achieved at some stage. (Moravec neatly summarises the arguments against several old chestnuts that claim such a thing is impossible.) This takes us to the second half of the book, where things get wilder, hairier, and much more interesting. Moravec discusses what happens to the robots, and what happens to us, when we are outstripped by our "mind-children". He describes ways in which the human race may survive, for a while at least, in the presence of transcendent intelligences, but claims that our eventual demise is not a tragedy, because these intelligences are, in a very real sense, our children, who just take over from their parents in the way of all children.

But even this might not be a problem. The final chapters get all eschatological, with another nice take on quantum parallel worlds and Gell-Man's "goblin worlds" ideas, leading to a startling conclusion about life after death.

So, I disbelieve the timetable for the short term predictions, because I think full embodied intelligence requires more than simply faster, cleverer classical algorithms. But this is definitely worth reading for those predictions and the reasoning behind them. And the second half is pure mind-expanding fun.