Why AI Failed -- The Past 10 Years
9 June 1996
(and a comment by Bill Gates)
There has been some progress in AI in the past decade, but
not much. I see five reasons for this:
- The field has shattered into subfields populated by
researchers with different goals and who speak very different technical languages. Much
has been learned, and it is time to start integrating what we've learned, but few
researchers are widely versed enough to do so. Marvin Minsky has a proposal for how to
do so in his seminal work, The Society of Mind, but his framework covers so much territory
that few have managed to understand it well enough to implement it.
- The field suffers from physics envy. Most AI researchers
are looking for simple explanations to inherently complex phenomena. Perhaps the universe
can be boiled down to a few simple rules, but it is unlikely that brains can. Consider
that roughly half of our DNA is dedicated to directing the growth of the brain and the
nervous system. To the extent that we ought to model ourselves after another science, we
should copy biology - there are doubtless hundreds of different kinds of mechanisms
in the brain, specialized to different sorts of tasks, integrated in an equally complex
management structure. The current excitement over neural networks, logical theories,
statistical approaches, and genetic algorithms is symptomatic of this puzzling, widespread
belief that intelligence can be captured by a simple mechanism.
- Many AI researchers have lost touch with the original goal
of building a flexible, human-level intelligence. This is partially a consequence of the
need to specialize in order to make a name for yourself in the field, and partially a
consequence of the funding structure of most institutions that pursue AI research. It is
difficult to get funding for risky, large-scale integration projects.
- AI is the ultimate software problem, yet many AI
researchers spend inordinate amounts of time building and maintaining robots. We should be
experimenting with new algorithms, not soldering! When robots are needed, we ought to work
in simulated worlds.
- AI researchers have been trying unsuccessfully to get
around the need for common sense knowledge. To solve the hard problems in AI - natural
language understanding, general vision, completely trustworthy speech and handwriting
recognition - we need systems with common sense knowledge and flexible ways to use it. The
trouble is that building such systems amounts to "solving AI". This notion is
difficult to accept, but it seems that we have no choice but to face it head on.
This prompted a brief note from Bill Gates:
"I think your observations about the AI field are
correct. As you are writing papers about your progress I would appreciate being sent
copies. I am still extremely interested in AI."
I'm glad someone is!